AI search news

On 2026-08-28 we shipped a month-sectioned /news index with next and previous links on every post, pinned the 14 posts Google had not indexed, and pre-registered a read for 2026-09-11: at least 7 of 14 indexed means the hub works. The read scored 8 of 14. Every indexed post has at least one Googlebot fetch on our server ledger since the hub shipped and Search Console's lastCrawlTime matches that first fetch to the second; the 6 posts with 0 fetches in 14 days are still out. The hub page itself was fetched 0 times by Googlebot in the last 7 days and 63 times by other crawlers. Tables, method, what the read cannot separate, and the next read.

Our cold campaign to US and UK web agencies opened every first email with a free AI-visibility report already run on one of the agency's own named clients. At its pause on 2026-09-11, read from the campaign's own sent list and PostHog: 97 agencies loaded, 92 contacted, 132 emails across 7 populations and 7 sending days, 5 bounced, 4 of 40 opened where opens were measurable, 0 loaded the report, 0 clicked, 0 replied. The full table, the method, what the give did not do, and the resume of 2026-09-18 with its read pre-registered.

On 2026-09-09 a reader in Lahore opened askedabout.com with no referrer at 1440×670, had the check form on the first screen, clicked the Pricing link, came back by the Run free check link, clicked into the Business name, What you do and City or market fields inside 3 seconds, clicked a line of hero text, scrolled to the bottom and stayed 580 visible seconds. 7 clicks, 4 of them flagged dead by PostHog's own detector, 1 focus event, 0 step-1 completions, 0 checks. The click-by-click table, the site's step funnel with its new typed column (live 2026-09-10), the same form walked at the same viewport, and the 2026-09-15 read, pre-registered.

A two-step cold campaign to US and UK web agencies runs under a daily limit of 20 across two inboxes with a 5-day wait before its one follow-up. Read on 2026-09-10 from the campaign's own sent list on Instantly: on all 6 completed sending days the number of first emails sent equalled the smaller of the rows available and the cap minus the follow-ups due (10/10, 0/20, 0/13, 0/19, 20/0, 10/10, follow-ups/first emails). On 2026-09-08 20 follow-ups filled the cap and 12 loaded rows waited a day. The per-day table, the capacity rule, the Monday stacking that a 5-day delay produces on a weekday-only window, and the forecast for the next three sending days, pre-registered.

Every session ChatGPT has sent to one askedabout.com post since 2026-08-08, read on 2026-09-09 from PostHog: 7 sessions, 7 person ids on 5 networks, 5 with the check-form call-to-action in view, 3 that scrolled to the form, 1 that focused a field, 0 that ran a check. The latest reader arrived from ChatGPT in Chrome, reopened the page in Edge 99 seconds later, stayed 2094 visible seconds, and never reached the form. The form moved to the claim on 2026-09-09; the 2026-09-15 read is pre-registered.

Our cold campaign to US and UK web agencies opens with a free AI-visibility report already run on one of the agency's own named clients, linked in the first email. On 2026-09-09 we read the campaign's own sent list against PostHog: 62 agencies had received that link 2 to 7 days earlier, 5 bounced, 4 of the 43 sent with open tracking opened, 0 loaded the report, 0 replied. Loads, opens, bounces and replies by population, the method, and the pre-registered read of 2026-09-11 that ramps or pauses the campaign.

Session-level read, 2026-09-08: every real session on askedabout.com over 41 hours (UTC 09-07 and 09-08 to 16:39 UTC) was 12 arrivals, 9 of them reads by 7 people, from Google (2), ChatGPT (2) and no referrer (5). All 9 landed on 3 posts published 21 to 30 days earlier. The 30 posts published since 2026-08-26 had 0 reads, 0 Google impressions and 0 clicks in Search Console over the same window, and 7 of the 7 newest are not indexed. Method, tables and artifact inside.

We ran every address that had received the first email of our cold campaign (62 named people at US and UK web agencies, all with an MX record) through Instantly's email-verification endpoint on 2026-09-08. 40 read verified, 13 read verified on a catch-all domain, 9 read invalid. Of the 5 addresses that had bounced, 1 read invalid and 4 read verified; of the 57 that did not bounce, 8 read invalid. Bounce rate by verification class, MX host and send day, the method, and a pre-registered read of the gate's first week.

Server-log read, 2026-09-07: PerplexityBot fetched 60 distinct paths on askedabout.com (33% of the 179 sitemap URLs) in two bursts, 44 rows in 9 minutes on 09-06 and 17 in 20 minutes on 09-07, its second whole-site sweep in 5.3 days. It reached 4 of the 5 newest posts 14.4 to 20.4 hours after publish. Googlebot logged 39 rows on 14 paths in the same 28.5 hours, 18 of them robots.txt, and none on the 5 newest posts. Perplexity's citations of us on the same day's gauge: 3 of 36, unchanged. Method, tables and artifact inside.

On the 2026-09-07 run of our 12-question, 4-engine, 3-sample panel, 87 of the 248 domains that entered the AI-cited set on 2026-08-31 were cited again (35.1%); 247 of 309 incumbents were (79.9%). Both sit inside the bands pre-registered last week, for the fifth transition running. Entrants introduced by ChatGPT survived most this week (49.4%), by Gemini least (24.6%); single-answer entrants 24.6% against 61.5% and 55.3% for two and three-plus answers. Method, script and artifact inside.

First-party read at 2026-09-06 16:46 UTC: a post published 2026-08-25 sat "Discovered" or "unknown to Google" through 9 daily URL Inspection reads with 0 Googlebot requests on its path, was fetched once by Googlebot on 2026-09-05 at 20:05 UTC (11.3 days after publish, 3.2 hours after a sitemap fetch), and read "Submitted and indexed" within 14 hours. On the same 14-URL index, the 2 posts Googlebot fetched are the 2 indexed ones; the 12 it never fetched are all still out, with 110 requests from other crawlers between them. Every request on the page's path, the fetch-day Googlebot log, the sitemap and IndexNow record, and the 2026-09-11 read pre-registered.

A 55-day read of one pitch campaign to named writers and editors at outlets AI engines cite, 2026-09-06 14:11 UTC: 107 emails, 66 people, 23 unique openers, 3 replies, 0 tracked link clicks. Opens, replies and clicks by first and second touch, opens per record, sends by week, and the read pre-registered for 2026-09-13.

First-party read at 2026-09-05 16:44 UTC: a data page on askedabout.com with 1 Google impression and 0 clicks in 28 days took 3 arrivals with no referrer in 10.7 hours, then 2 ChatGPT-User fetches 17.9 and 21.3 hours after the first visit. One iPhone reader spent 6 min 28 s on the page over two sessions, scrolled to 100% and 83%, saw the check form three times and never focused it. What a referrer-less iOS arrival can and cannot be attributed to, with the fetch ledger, the analytics rows and the Search Console record.

Third read of a give-first cold-email restart to named people at 43 US and UK web agencies, 2026-09-05 15:46 UTC: 53 emails, 3 unique openers, all on non-Google mail and all contacted before a sequence fix; 0 of 21 Google-hosted recipients opened. The one variable changed today, the pre-registered read on 2026-09-09, and the counts by provider, population and hour.

Google put Gemini 3.8 Flash into AI Mode for AI Pro and Ultra subscribers on 2026-09-02. Within a day, SEOs posted AI Mode answers on that model with no links or citations at all. Robby Stein, VP of Product for Google Search, replied on X: "This isn't working as intended, and we'll roll out a fix soon." Search Engine Land reports the fix landed the following morning and that the bug only affected the paid-subscriber model. What it means for anyone measuring AI Mode visibility.

Search Console read at 2026-09-04 16:44 UTC: 6 Google web-search clicks on askedabout.com in the 8 landed days of a 28-day window whose bar is 25, a pace of 0.75 a day against 0.95 needed on the 20 days left. 5 of the 6 clicks are on one post published 2026-08-11, 1 on a post from 2026-07-28. The 21 posts published inside the window have 0 impressions between them; 8 of 8 inspected today read "Discovered, currently not indexed". Per-day series, per-page table, PostHog cross-check and the 2026-09-23 read, pre-registered.

Second read of a give-first cold-email restart to named people at 31 US and UK web agencies, from the Instantly API and PostHog at 2026-09-04 13:52 UTC, 47.9 hours after the first send. 41 emails, 3 unique openers (10%), 0 link clicks, 0 loads of the linked free reports, 0 replies, 0 bounces. Split by population: the 10 leads contacted before the 2026-09-03 sequence fix (two emails each, 9 minutes apart) hold all 3 openers; the 21 contacted after it (one email each) show 0 opens at a median of 23 h. Timing series, per-lead table, and the pre-registered reads for 2026-09-05, 09-08 and 09-09 inside.

Google's 2026-09-02 launch post for Gemini 3.8 Flash names AI Mode in Google Search directly: the model is available to Google AI Pro and Ultra subscribers across the Gemini app, AI Mode and Sheets. It is the third Flash release in six weeks, three weeks after 3.7 Flash. What the post says, what it does not say about free-tier AI Mode or AI Overviews, and why a dated, tiered read is the only kind that counts.

Our /news index shipped on 2026-08-28 as an internal-linking fix for 14 blog posts Google had discovered but not crawled. Googlebot fetched the index on 2026-09-02 at 14:49:18 UTC, 58 days after its previous crawl, and twice more 19.6 hours later. Seven daily reads of the 14 linked posts: 1 indexed, 12 Discovered, 1 unknown, 45 state flips, 13 posts with 0 Googlebot fetches in 7 days. PerplexityBot fetched 8 of them, Amazonbot 13. Two-channel evidence, per-URL table, and the 2026-09-11 read pre-registered.

We restarted cold email on 2026-09-02 with a give-first hook: a free AI-visibility check on one of the agency's own named clients, sent to a named person at 16 US and UK agencies. Read from the Instantly API and PostHog at 24 hours: 26 emails, 3 unique openers of 16 (19%), 0 link clicks, 0 loads of the free reports, 0 replies, 0 bounces. A sequence-configuration defect sent all 10 first-day recipients their one follow-up 9 min 01 s to 9 min 04 s after the first email. Per-lead table, the cause, the fix, and three dated reads inside.

Eight days into a 28-day window we pre-stated on 2026-08-19, our AI-referral sensor reads 11 gross and 7 net referrals from ChatGPT (5), Gemini (1) and Claude (1) against a bar of 3. Lifetime since 2026-06-27: 19 net from four engines, ChatGPT 15. What the seven did: every one fired an engagement tick, median 38 visible seconds, 2 saw the form, 2 focused a field, 1 ran a check, 0 upgraded. The four held-out rows are one fixed-window shape. Tables, method and the next dated read.

On 2026-08-26 we published three falsifiable readings about where AI crawlers go on a small site and dated the re-read for today. Week 3 (2026-08-26 13:37 to 2026-09-02 13:30 UTC, 2278 net bot requests): ChatGPT-User put 219 of 309 requests on the homepage (71%; bar 50%, falsified under 30%), Googlebot put 109 of 269 on robots.txt and sitemap.xml (41%; bar 40%), Amazonbot touched 166 distinct content pages against 69 for the widest AI fetcher. Three-week table, method, and the next dated read inside.

Every /news post we published 2026-08-24 to 08-31 was fetched by ClaudeBot (median 1.7 h), OAI-SearchBot (4.7 h), PerplexityBot and YandexBot. Googlebot fetched none of the 17 at 23 h to 8 d, while making 265 requests elsewhere on the site in the same 7 days, 94 for robots.txt. Ledger rows, per post, per crawler.

In the 28 days to 2026-08-30, 45 askedabout.com URLs averaged Google position 10.0 or better: 804 impressions, 11 clicks, 1.37%. Seven clicks came from one URL that moved from 10.2 to 9.2; the other 44 read 0.53%. Six windows of Search Console page rows, computed by script.

Google updated its June announcement on August 31, 2026: the Search Console control that lets site owners opt out of AI Overviews, AI Mode and AI Overviews in Discover, and the AI performance report, have 'rolled out to all websites worldwide.' Opting out is not a ranking signal elsewhere, takes effect within days, does not affect AI training, and costs every impression from Google's generative AI features. What the toggle does, what it does not, and what it means for measuring AI visibility.

Our server-side crawler ledger counted 206 ChatGPT-User fetches of askedabout.com's homepage in the 7 days ending 2026-08-31 13:11 UTC, up from 148 the week before. In the same weekly gauge run, 3 of 36 sampled OpenAI answers cited the homepage, down from 5 of 36. Three weeks of the pair, per question cell, with the script and artifact committed. Fetch counts in your logs measure retrieval, not recommendation.

On the 2026-08-31 run of our 12-question, 4-engine, 3-sample panel, 85 of the 256 domains that entered the AI-cited set on 2026-08-24 were cited again (33.2%); 224 of 303 incumbents were (73.9%). Both sit inside the bands pre-registered last week. Entrants introduced by Gemini survived least (22.9%), by Perplexity most (66.7%); single-answer entrants 19.7% vs 73.9% for three-plus. Method, script and artifact included.

In the trailing 30 days, 7 sessions on askedabout.com arrived with a browser viewport of exactly 1280×680 on four different screen sizes, fired 1 to 4 events, recorded 0 engagement ticks, 0 scroll events and 0 leave beacons, and were gone in under 2 seconds. Six carry utm_source=chatgpt.com; five were counted as AI referrals by our own detector. None of our 14 other eligible AI-referral sessions has that viewport. We now hold the shape out of our own Day-103 checkpoint: 5 gross, 2 net. Every session as a table.

We joined 7 days of our first-party crawler ledger (2,146 bot requests, 240 from Googlebot) with Search Console's page-one URL set (46 URLs at position 10 or better, 840 impressions, 11 clicks). Googlebot re-fetched 16 of the 46. The 30 it did not touch carry 622 of the 840 impressions, including the three largest (181, 108 and 87 impressions, 0 fetches). 0 of the 14 posts published in the last 7 days were fetched. Method, table and a dated re-read inside.

Every AI-engine referral to askedabout.com in 78 days, as counts: 17 arrived, 7 had the check form in view, 3 tapped the button, 2 focused a field, 1 started a check, 1 read a 0/100 report twice, 1 visited pricing unprompted, 0 clicked upgrade. Both form touches came from AI engines (claude.ai, then ChatGPT) onto news posts, 13.3 hours apart. Event rows, timestamps and what one session cannot separate.

Google Search Console's per-day web rows for askedabout.com, 2026-07-30 to 2026-08-26: 2,884 impressions, 13 clicks, 19 zero-click days. The seven highest-impression days (1,090 impressions) produced 1 click; the seven lowest (377) produced 2. Spearman rank correlation between daily impressions and daily clicks: -0.06. Google disclosed a query for 0 of the 13 clicks. Every daily row, the streaks, and a pre-committed re-read on 2026-09-26.

The pre-registered 72-hour read of our Request-indexing test, as counts. Of 9 unindexed URLs nobody clicked, 2 were indexed on 2026-08-27 from one organic Googlebot fetch each, during a day when Googlebot's requests to the site doubled (53 rows, 32 paths, vs 27 and 15). All 10 clicked-and-indexed URLs were re-fetched organically within 72 hours, yet Search Console still reports the click-second as their last crawl. The 13 URLs Googlebot has never requested, plus the one click that never fetched, are 0 of 14 indexed with 0 Googlebot rows in 7 days. Per-URL tables, what one day cannot separate, and the internal-linking variable we shipped today with its 2026-09-11 read.

Every session our Uneed listing (live 2026-06-26) and Product Hunt launch (2026-06-30) ever sent to askedabout.com, read back from PostHog and classified by our standing instrument: 30 sessions, 29 real, 22 network addresses, 8 from one Lagos household. 2 of the 4 sessions where the form was measurable saw it; 1 session, on Product Hunt launch night, started a check; 0 bought. Method, the full 30-row ledger, and the read pre-committed for 2026-09-28.

Fifteen visitors have arrived at askedabout.com from a ChatGPT or Perplexity answer since launch. Sessions ran from 0.5 seconds to 30.5 minutes, median 29.8 seconds; four ended inside two seconds and all four landed on a product page, while no article landing ended inside six seconds. One visitor tapped the check button, none typed into a field, none ran a check. The newest referral, on 2026-08-27 at 09:04 UTC, sent four events in 0.9 seconds and nothing more in the next seven hours. Every session timestamped.

We read back every email our Instantly campaigns ever sent (306 sends, 8 campaigns) straight from the API and classified the first line. 96 of 97 citation pitches opened "Hi ," because the template asked for a variable the lead records did not carry; 61 of 191 cold emails did the same. Both human replies came through a blank greeting. Method, per-campaign table, the 50-day timeline and the pre-committed Monday read.

On 2026-08-25 a scheduled task clicked Request indexing in Search Console on 11 URLs that Google had left unindexed for 5 to 39 days. Our crawler ledger saw a Googlebot fetch on 10 of them within about two minutes of each click, in click order. 27 hours later all 10 are indexed and Search Console reports that exact fetch second as the last crawl. The 11th never fetched and is still out. Nine unindexed URLs we did not click: 0 indexed, 0 Googlebot visits in 9 days.

Seven days of our first-party crawler ledger (1,955 bot requests after excluding 58 probes and 22 spoofed scanner rows), cut by fetcher and path class. ChatGPT-User fetched the homepage 193 times (71% of its 272); Googlebot spent 56% of 176 requests on robots.txt and sitemap.xml; Amazonbot touched 151 distinct content pages, more than any AI fetcher. Method, tables and the pre-committed 09-02 read.

Across 74 days, five external visitors reached the free-check form on askedabout.com. The three who clicked into a field all submitted and saw a report within 40 to 94 seconds of landing; one reached checkout. The two who only tapped the call-to-action button, one from Google on a 2560-px desktop, one from ChatGPT on a 390-px iPhone, never touched a field. Of 14 AI-referred visitors, one tapped, none typed, none ran a check. Every session timestamped, with the pre-08-04 sensor gap stated.

We asked ChatGPT, Perplexity, Gemini and Claude the same 12 buyer questions three times each on four Mondays (576 production answers). askedabout.com was in the returned citation list 18 times: 9 from one ranked-list page Perplexity cited 3 of 3 on three straight runs, 9 from the homepage on ChatGPT. Gemini and Claude 0 of 144 each; the two pricing questions 0 of 96; the same-shaped control question with only a head-to-head page 0 of 48. Every cell, the ranks, and the 08-31 falsification.

A post we published at 16:49 UTC on August 23 was fetched by OAI-SearchBot 61 minutes later and received a chatgpt.com-referred human reader 11.2 hours after publish, who stayed 28.5 minutes. Googlebot had not fetched the page then and still had not at 24 hours; Search Console reports it as unknown. Full timeline from three instruments, the four sibling posts, and the pre-committed August 28 read.

Five weekly runs of 12 buyer questions × 4 engines × 3 samples. A domain newly cited one week is cited again the next only 27.5–30.1% of the time; a domain cited two weeks running returns 75–78%. Per-engine, per-breadth and per-answer-count tables, the method, and the pre-committed 08-31 read.

Google Search Central: JSON-LD extraction now applies a single pass of HTML unescaping and double-escaped entities are no longer unrolled. We parsed the structured data on 572 pages AI engines cited for 12 buyer questions: 443 carry JSON-LD, 79 (17.8%) ship entity text in a string value or fail a strict parse, 63 hosts including ahrefs.com, blog.hubspot.com, techradar.com, searchengineland.com and developers.google.com. 56 have it in a headline, name or description. Method, both readings of Google's rule, and the full table inside.

Search Console discloses 265 queries for askedabout.com over 28 days (1049 impressions, 0 clicks). 83 of them (31.3%) are prompt-shaped: role-prefixed first-person paragraphs, 77-word vendor briefs, and 10 queries carrying a literal <context> block with another company's URL. They average position 72.4 and earned no clicks. All 12 of the property's clicks came from the 1523 impressions Google attributes to no disclosed query. Method, rules and the full 265-row table inside.

At 18:19:56 UTC on 2026-08-21 a scanner in Singapore made 22 requests to this site in 4.5 seconds — for /.env, /.git/HEAD, /.aws/credentials, /etc/passwd and similar — while rotating six AI-crawler user-agent strings: GPTBot, ChatGPT-User, OAI-SearchBot, Google-Extended, Amazonbot and Amzn-SearchBot. Our user-agent-classified crawler ledger booked every one as a real vendor fetch. Net of the scanner, this week's GPTBot count falls from 11 to 7, Amzn-SearchBot from 6 to 2, and Google-Extended from 4 to 0 — every Google-Extended row the ledger has ever held was the scanner, and Google's documentation says no such user-agent string exists. The request table, the vendor verification docs, the gross-vs-net counts, and what we changed.

Since 2026-08-04 we have read the same Search Console bucket every day: every page on this site whose trailing-28-day average position is 10 or better. Impressions in the bucket went from 383 to 881 (+130%) and the URL count from 30 to 51 (+70%); clicks read exactly 4 on ten of the nineteen days and never cleared 5 on a clean read. CTR 0.45%. 47 of the 51 page-one URLs earned 0 clicks on 753 impressions, the two largest (142 and 105 impressions at positions 7.9 and 7.8) among them. Of 255 queries Google discloses, exactly one averages a page-one position — and it is a ChatGPT citation-artifact string. The per-read series, the per-URL table, the four titles we rewrote six days ago, and the date the read is decided.

Every post this site published between August 18 and 20 — eight pages — has zero Googlebot fetches on our server-side crawler ledger at ages 24 to 75 hours. In the same window OAI-SearchBot fetched all 8 within 12 hours of publish, PerplexityBot all 8 within 18, and Yandex or Bing reached 7 of 8 within 3 minutes. Googlebot is not absent: it made 122 requests to the site in 4.6 days, including at least 9 page fetches today — every one of them on a page published August 17 or earlier. We publish the per-post ledger, the Search Console states (three now read 'URL is unknown to Google'), and the control: the one post Googlebot did fetch this morning was indexed the same day, after 8 days stuck.

Our AI-referral detector has fired 10 times in its lifetime: 8 ChatGPT referrals and 2 Perplexity referrals across 70 days of a small site, every fire timestamped with its landing page. The series took 34 days to produce its first six fires, then produced four more in 28.6 hours. The same 8 landing pages earned 6 Google clicks in the trailing 28 days — and two of the ten AI referrals landed on a page Google ranks at position 75.6 with 0 clicks on 126 impressions. We publish the full series, the two referrals our own analytics lost before a fix, and the number that cuts against the story: 0 of the 10 sessions has started a check.

Two crawlers our server-side ledger had never recorded showed up in the same week: IbouBot, the crawler of Ibou.io, a French search engine whose documentation says every answer it produces cites and links its sources, and PetalBot, Huawei's crawler, whose documentation says it feeds Petal Search plus Huawei Assistant and AI Search. Between 2026-08-17 and 08-20 they made 136 fetches of one small site — IbouBot swept 46 distinct paths in one 12-minute pass, PetalBot accelerated from 8 fetches a day to 59. What each bot is, from its own documentation, what it read (both fetched robots.txt; neither touched llms.txt), and whether you should block them.

Overnight on 2026-08-20 our analytics recorded three arrivals in 92 minutes on one comparison page — the first ChatGPT referrals ever onto that section of our site. Google Search Console shows the same page at average position 75.8 with 128 impressions and 0 clicks in 28 days. The first visitor arrived 3 minutes 11 seconds after OpenAI's ChatGPT-User agent fetched the exact page. None of the three sessions recorded any event after the landing pageview, and one of the three referrals was lost by our own client-side event batching — both facts we report beside the finding. First-party session rows, crawler ledger and Search Console data, all timestamped.

We timed every crawler's first fetch of the seven posts we published between 2026-08-17 and 08-19, from our own server-side crawler ledger. YandexBot arrived within 1–25 minutes of every publish (IndexNow), OAI-SearchBot reached all six posts older than a day in 22 minutes to 5.6 hours (median 5.3 h), PerplexityBot four of six in 7–30 hours, Amazonbot five of six in 13–27 hours. Googlebot read the resubmitted sitemap within two minutes every time but fetched only three of the six pages, at 11, 15 and 34 hours; the other three are 'Discovered – currently not indexed' at 24–27 hours. GPTBot and ClaudeBot fetched none. One ChatGPT-User fetch arrived at 39 hours.

Promptwatch reports reddit.com's share of ChatGPT Search citations fell from a 3.83% average (July 18–August 7) to 0.52% over August 14–17, an 86.4% drop, with a smaller step on August 8; Google AI Overviews and AI Mode declined gradually (−11.3%, −30.5%). Promptwatch itself says a data-collection issue cannot be ruled out. Our own weekly read of the same engine through OpenAI's search API, three days after the cliff, found reddit.com in 9 of 36 ChatGPT answers on August 17 — the highest of four weeks (2, 8, 2, 9) — while Perplexity cited it in 28 of 36. Surface matters: the consumer ChatGPT Search interface and the API are not the same pipe.

We asked how many of 12 closely related small-business AI-visibility questions each cited domain appears on, across four dated weekly runs of ChatGPT, Perplexity, Gemini and Claude (576 production answers, 5,949 domain-answer slots). In every run 78.0–79.9% of cited domains appeared on exactly one question; only five or six per run spanned six or more; the ten most-cited domains took 13.4–15.2% of slots. A domain cited on two or more questions in one week was cited again the next week 87.0% of the time (295 of 339) against 46.0% for single-question domains (579 of 1,259). reddit.com spanned 11 of 12 questions in all four runs.

Google's Search Console data-anomalies page now records a logging error that lowered impressions on the Generative AI performance report in Search from August 13, 2026 onward — 'ongoing' as of August 17 — plus a one-day Discover error the same day. John Mueller: 'just a logging issue and not representative of visibility changes in Search'; an annotation is coming. We read our own property's day-by-day web totals across the break: no step down (113.5 impressions/day after August 13 vs 105 before), consistent with the bug being confined to the AI report's logging. What agencies should and should not tell clients this week.

Dan Petrovic (DEJAN) published a server-log census of dejan.ai's three machine files for the 30 days to 2026-08-15: Google, Anthropic and OpenAI crawlers fetched robots.txt 1,150 times, the OKF content bundle 23 times and llms.txt 4 times — none from OpenAI; the three Googlebot llms.txt fetches were followed by no page request. Of 157 llms.txt fetches from all agents, 54% carried a browser user agent and 8% were curl. We read our own crawler-hit ledger over its first 39 hours: robots.txt fetched 65 times by bot-shaped agents (Googlebot 16, OAI-SearchBot 10, PerplexityBot 1, Claude-User 1), llms.txt once, by no recognised AI crawler. Google's John Mueller, June 2025: 'no AI system currently uses llms.txt.'

We counted distinct normalized URLs per cited domain across four dated runs of 12 small-business buyer questions on ChatGPT, Perplexity and Claude (576 production answers, 4,505 path-carrying citation slots). Of 713 cited domains, 81.6% contributed exactly one URL, 18.4% two or more, 8.1% three or more; mean 1.53 URLs per domain. The multi-URL minority is the stable minority: domains cited via ≥2 URLs in one week were cited again the next week 92.7% of the time (178 of 192) against 55.0% for single-URL domains (482 of 877). Gemini excluded — its channel returns bare hostnames.

Google Research (blog 2026-08-12; paper arXiv 2602.14080, 'Empty Shelves or Lost Keys?') profiles 2,150 Wikipedia facts across 13 LLMs and finds the bottleneck is recall, not knowledge: Gemini-3-Pro and GPT-5 encode 95–98% of facts yet fail to directly recall 26–34%, and still miss 11–12% with thinking on. Rare facts are encoded at nearly the rate of popular ones but recalled far less; reverse-direction questions ('What is A?' when the model learned 'A is B') are recognised as easily as direct ones in multiple choice but generated far less often — 'the reversal curse is a recall problem.' Thinking recovers 40–65% of encoded-but-unrecalled facts. That is the shape of our own July data: engines named 168 local businesses in 98.8% of by-name answers and 19.8% of buyer-question answers.

RESONEO's reverse-engineering study of 1,249 real ChatGPT answers (July 2026, updated August) finds ChatGPT rarely reads the page it cites: it works from a title, a URL and a ~202-character snippet fetched from one of four engines — OpenAI's own 'labrador' index or scraped Google (Bright Data / Oxylabs). Only 759 pages were opened across 61,332 side-panel URLs (1.2%), 757 of them in paid thinking mode; an opened page is cited 74% of the time, a merely-retrieved page 7%. Free instant mode is 63.6% labrador; paid thinking is 74.6% scraped Google. The labrador snippet is cut at 202 characters from the start of the body, anchored on the H1, meta description ignored. In our own 144-answer run this morning, ChatGPT cited a homepage in 26.5% of its 279 citations (Perplexity 1.1%, Claude 1.9%, Gemini 0%) — all four of our own citations from ChatGPT were the homepage.

We asked ChatGPT, Perplexity, Gemini and Claude the exact question — "What tools track whether AI recommends my business?" — three times each, every week for four weeks (48 production answers), and tallied which tools the answers name and which pages they cite. Otterly.ai is named in 28 of 48 answers (58.3%), Profound in 27 (56.3%), Peec AI in 24 (50.0%), Sight AI in 23, Semrush's AI Visibility Toolkit in 21; 42 tools were named at least once and 15 of the 211 cited domains appear in all four weeks. The most-cited page is not a vendor's product page but Sight AI's own listicle (34 of 48 answers). AskedAbout is named in 0 of 48. Entry prices as read from each vendor's own pricing page on 2026-08-01, machine-readable table included.

TollBit's State of the Bots report for H1 2026 (published this week) finds ChatGPT-User — OpenAI's live page-fetcher — ignored an explicit robots.txt disallow on 54% of its scrapes across European publisher sites, the highest bypass rate of any AI bot measured, while averaging ~3.16M scrapes per site. OpenAI's own crawler documentation says exactly this can happen: because ChatGPT-User fetches are user-initiated, "robots.txt rules may not apply." We re-ran our guarded robots.txt parser on the 150 domains AI engines cite most on our 12 small-business buyer questions: 142 measurable, ZERO fully disallow ChatGPT-User, only 38 name it at all, and only 2 fully block GPTBot — the sites that get cited are the sites that do not block.

We tokenized the URL path of every deep page four AI engines cited across three dated runs of 12 small-business buyer questions (432 production answers, 3,030 deep-page citations). 91.8% carry at least one word from the question asked in the path, 77.1% carry two or more, and 82.2% carry a distinctive question word once generic vertical tokens are excluded. ChatGPT is the least slug-literal engine at 77.2% (Perplexity 95.1%, Claude 94.4%). Only 5.1% of cited paths carry a year — 152 of the 154 that do say 2026, and not one carries a year older than 2025.

Google introduced Gemini 3.7 Flash on 2026-08-13 and by 2026-08-14 it was rolling out in Search's AI Mode as a selectable model for Google AI Pro and Ultra subscribers, globally in English — picked from the "+" menu alongside "Auto" and "Pro." The measurement consequence: "how does Google AI Mode describe my business" no longer has a single answer. The generator now varies by subscription tier and by which model the searcher selected, and most AI-visibility tools measure neither the paid tier nor the menu.

We normalized every URL four AI engines returned across three dated runs of 12 small-business buyer questions (432 production answers, 3,281 path-carrying cited URLs). Only 7.7% pointed at a homepage — 92.3% cited a deep page, and the single most common shape is a two-segment path like /blog/answer-slug (61.3% of cited URLs). ChatGPT is the outlier at 28.4% homepage citations; Perplexity reads 1.7% and Claude 0.3%. Gemini's citation channel returns bare hostnames and cannot testify either way.

On 2026-08-13 Microsoft added an AI Scrape-to-Referral Ratio to Clarity's Bot Analytics: one number comparing how often each AI operator scrapes your site against how many visitors it sends back, with a ranked per-operator breakdown and session recordings of the AI-referred visits. The catch: it reads server-side logs, so it needs a CDN or server integration — the free tracking snippet alone cannot see bots. And a scrape-to-referral ratio undercounts AI's value for the same reason referral rows do: most AI-shaped visits arrive with no referrer at all.

Shopify's analysis of Q2 2026 storefront data (published 2026-08-11): AI-referred sessions grew 197% year-over-year and AI-referred orders roughly 3x, while organic search sessions grew 12% on a much larger base. AI-referred shoppers who reached product pages converted about 80% better than organic overall — roughly 2x in specification-heavy categories — and half of AI-referred sessions land directly on product detail pages. Stores exposing structured product data converted AI visitors at 2x the rate of stores the engines had to scrape.

We asked four AI engines 12 identical small-business buyer questions three times each, in three separate weekly runs (432 production answers, 2,623 cited-domain slots). Only 25.4% of cited domains appeared in all three answers to the same question asked minutes apart — 57.0% appeared exactly once. The spread by engine is 9.7×: Perplexity keeps 63.8% of its cited domains across all three asks, Claude 23.9%, ChatGPT 7.4%, Gemini 6.6%. A single screenshot of a ChatGPT answer is not a measurement.

Kevin Indig's analysis for G2 (~35,000 ChatGPT citation URLs, December 2025, 3,177 SaaS prompts) finds UGC platforms hold 17.1% of cited domains — 4x publishers at 4.0% — with Wikipedia, Reddit and LinkedIn accounting for 99% of UGC citations. We ran the same class of cut on our own three four-engine small-business buyer-question runs: the same three platforms hold just 3.2–4.0% of domain-answer citation pairs, Reddit carries most of it, and Wikipedia is nearly absent. UGC share is a property of the vertical and engine mix, not a constant.

A Scrunch panel study published 2026-08-13 (millions of search events, February–June 2026) finds readers are 20.5 percentage points more likely to visit a news site in the week after a news conversation with an AI assistant — but only 1.1% of those visits arrive with an AI referrer. About three quarters show up as direct navigation. The study also finds the famous AI Overviews click gap (20% vs 30%) shrinks to about 2 points when the same query is compared with and without an overlay.

Per-question cut of three identical weekly four-engine runs (12 buyer questions × 4 engines × 3 samples = 144 answers/run): a single question's citation field ranges from 35 to 93 distinct domains per run, and width is a stable property of the question (rank correlation 0.69–0.79 across runs). Narrow questions keep their citation sets — the three narrowest questions hold 32.3% of their domains across all three runs, versus 14.4% for the three widest, a 2.3× persistence gap.

Google filed its amended complaint against SERP-scraping vendor SerpApi on 2026-08-10, three weeks after Judge Yvonne Gonzalez Rogers dismissed the original DMCA case — ruling that circumventing Google's SearchGuard is not a DMCA violation where the results contain no copyrighted content, and dismissing that portion without leave to amend. The rebuilt case turns on licensed content: Google quotes its Reddit agreement directing it "not to enable third parties to extract and independently commercialize the licensed content." Why the scraped-SERP supply chain behind most AI-visibility tools should watch this docket.

SE Ranking analyzed 50,006 commercial prompts across 20 US niches: ChatGPT now shows a sponsored slot on 25.94% of them — but the advertiser's domain is cited in the answer above the ad in only 3.63% of placements, and the exact advertised URL in 0.09%. We checked the study's ten heaviest ChatGPT advertisers against our own three-run, 969-domain organic citation corpus: nine never appear at all.

Three identical weekly runs of our four-engine citation corpus (12 buyer questions × 4 engines × 3 samples) let us finally ask what churn means: of the 249 domains that vanished between 2026-07-28 and 2026-08-03, only 47 (18.9%) were cited again on 2026-08-10 — 81.1% stayed gone. But the return rate splits hard on depth: domains that had been cited in one answer returned at 11.8%; domains cited in 2+ answers returned at 50.0%. Across all three runs, 969 distinct domains were ever cited, 58% in only one week, 23% in all three.

Search Console's Generative AI performance report is now showing on far more properties — Search Engine Roundtable's Barry Schwartz found it on all ~25 he checked on August 11, 2026, though Google's help doc still says 'subset.' We tested the obvious follow-up on our own property: can an agency pull that data through the Search Analytics API? Six candidate type values, six 400 INVALID_ARGUMENT responses; Google's API reference lists exactly six supported types and none is generative AI. The report is UI-and-export-only — and its impressions are already blended into your API 'web' totals with no way to split them out.

On 2026-08-10 Perplexity cited askedabout.com at position 3 of 20 sources in 3 of 3 identical samples of one buyer question — the first non-zero read in our weekly citation gauge (run 1: 0 of 144). We dumped every citation slot around it: 60 slots, 25 distinct pages, 21 of them multi-vendor alternatives lists, 2 Reddit threads, 1 directory page, 1 single-vendor pitch — and zero head-to-head comparisons. The cited page was built 2026-07-31 to test exactly that shape hypothesis. Survival read pre-registered for 2026-08-17.

Yelp announced on August 10, 2026 that ChatGPT users can book a table or join a waitlist at thousands of restaurants across the U.S. and Canada through Yelp, without leaving the chat. Participation runs through Yelp Guest Manager — a product roster, not a ranking anyone can optimize into. Our pre-deal baseline still frames the citation half: 0 of 511 ChatGPT local answers cited yelp.com before the licensing deal.

Second consecutive weekly re-run of our four-engine citation corpus: 227 of the 520 domains cited on 2026-08-03 were gone by 2026-08-10 (43.7%, vs 48.5% over the prior six days), and 247 new domains entered. The multi-engine protection effect replicated: domains cited by ≥2 engines survived the week at 87.0% vs 49.0% for single-engine domains. Only 225 domains have appeared in all three dated runs.

We counted the citations in every answer of our 2026-08-03 four-engine run (12 buyer questions × 4 engines × 3 samples = 144 production answers). Perplexity returned exactly 20 citations in 34 of its 36 answers. ChatGPT averaged 7 (range 4–11). Gemini averaged 12. Claude cited nothing at all in 16 of 36 answers — 44%. A 'citation' means a different thing on every engine, and the denominator is the part nobody quotes.

We joined our two dated citation cuts: which domains survived the six days between our 2026-07-28 and 2026-08-03 four-engine runs, and how stable each citation was across three identical asks within the later run. Incumbent domains held all three samples 36.3% of the time (219 of 603 citations); brand-new entrants only 12.4% (32 of 258). The gap holds inside every engine — ChatGPT gave new entrants 0 stable citations in 70, Claude 0 in 19.

OpenAI's August 6 Signals release headlines a shift "from asking to doing." We downloaded the public CSVs behind it. June 2026: Practical Guidance 32.1% and Seeking information 18.8% — question-shaped messages are 50.9% of all consumer ChatGPT use, and Seeking information (the category that includes purchasable-products queries) ROSE from 16.7% two years ago. Outside work, asking (45.1%) is still double doing (22.0%).

We classified every full citation URL in our 2026-08-03 four-engine run (12 small-business buyer questions × 3 samples) by path: homepage or deep page. Of 1,147 URLs from ChatGPT, Perplexity and Claude, 92.6% pointed below the homepage. Claude cited zero homepages in 178 citations; Perplexity 17 of 717 (2.4%) — all seven domains AI-visibility vendors. ChatGPT is the outlier at 27%, and its homepage citations concentrate on tool-discovery questions (74–77%) while how-to questions got zero.

Google's Search Console methodology doc, confirmed by John Mueller this week, says AI Mode data lands in the general performance report — and every follow-up turn inside an AI Mode conversation is logged as a brand-new query. That is why site owners are finding queries like “yes,” “yes go on,” and “yes, pricing” in their query tables. The Generative AI report has no query data; the report with query data has no AI label.

We ran a query-string census of every citation URL in our 2026-08-03 four-engine run (12 buyer questions × 4 engines × 3 samples, 144 answers, 1,587 cited URLs). ChatGPT appended utm_source=openai to 252 of 252 citation URLs — 100%, across 150 distinct sites. Gemini (440 URLs), Claude (178) and Perplexity (717) appended tracking parameters to zero. Traffic from three of the four engines arrives unmarked — AI referral counts built on link tags are a ChatGPT-only floor, not a total.

Cloudflare announced Cloudflare Wallets and cloudflare.pay on 2026-08-04: human-funded Account Wallets that delegate spending to agent-run Virtual Wallets, paying over the x402 protocol with stablecoins, under owner-set allowances, allowlists and per-transaction caps. Handles are claimable now; the wallet infrastructure ships over the coming months. For businesses, the AI recommendation and the purchase are converging into one machine-to-machine step.

Kevin Indig analyzed Semrush's US ChatGPT visibility data — 1,094 categories, 5 prompt variants each, 283,215 citations and 76,493 named-brand mentions from January–June 2026. In categories close to a brand's expertise, 74% of appearances are citations, 44% are named mentions, 34% get both; in distant categories mentions fall to 25%. Depth across prompt variants lifts mention share from negative to positive — except in legal and healthcare, where it stays negative even at full coverage.

We asked each of 12 buyer questions three times per engine (144 answers, 2026-08-03) and scored every cited domain by how many of the three identical samples it appeared in. 53.9% of 854 (question, engine, domain) citations appeared in only one sample; 28.9% in all three. ChatGPT and Gemini keep just 8% of their citations stable across three asks; Perplexity keeps 76%. Claude returned zero citations in 16 of 36 samples.

The Verge reports a wave of AI-search (GEO) spam on Reddit: agencies claiming to have “hijacked” subreddits, moderators banning brands, and Reddit blocking 23 million spam views a day with LLM-based detection. Google went on record that Reddit “gets no special preference” in Search or its AI features. Our own four-engine data: reddit.com is the single most-cited domain in 144 buyer-question answers — 46 answers cite it — but 32 of those are Perplexity's. Claude cited it zero times in 72 answers across two runs.

We diffed the multi-engine membership of our four-engine citation corpus across six days (2026-07-28 vs 2026-08-03). The core of domains cited by ≥2 AI engines held its size exactly — 99 domains both runs — but swapped a third of its members: 34 domains entered (26 climbed from single-engine, 8 were previously cited nowhere) while 27 were demoted to one engine and 7 vanished. Multi-engine citations are durable once held — and the door in is open: 5.9% of the single-engine tail climbed in one week.

Microsoft's Clarity Citations dashboard now labels every AI grounding query as branded or non-branded, breaks Share of Authority down by the two, and lets you filter the whole dashboard by query type — free, first-party data on the one split that matters: does AI cite you only when the user already typed your name, or does it choose you during generic discovery?

We asked ChatGPT, Perplexity, Gemini and Claude the same 12 small-business buyer questions twice, six days apart — 144 answers each run. 48.5% of the 538 domains cited in the first run were cited nowhere in the second. But survival splits cleanly on engine breadth: domains cited by one engine survived at 42.1%, by two at 90.2%, by three at 96.4%, and all 10 four-engine domains survived. A single AI citation is a lottery ticket; a multi-engine citation is an asset.

BrightLocal's 2026 Local Consumer Review Survey (n=1,002) found 45% of US consumers have asked an AI tool for a local business recommendation — up from 6% a year earlier. We paired that demand-side number with the supply side from our own audit corpus, re-measured today: across 179 local businesses and 707 unbranded buyer answers, the audited business is named in 14.9% of answers, and 67% of businesses appear in none of theirs.

We fetched 8 conventional contact pages on every one of the 439 domains that exactly one AI engine cited on small-business buyer questions. 12.8% were unreachable over HTTPS, 31.0% publish no email address at all, 46.5% publish only role aliases (info@, support@, privacy@), and 41 domains — 9.3% — publish a named person's email anywhere we looked. If your AI-visibility plan is “get cited by the sources the engines cite,” 9 in 10 of those sources offer no named human to talk to.

"60-70% of ChatGPT's local results come from Foursquare" is the most-repeated piece of local-AEO advice of 2026. We traced it: every page making the claim sources it to one 2025 measurement that none of them links, and neither OpenAI's nor Foursquare's own site announces the relationship. Then we checked our own corpus — 526 ChatGPT answers recommending local businesses, 1,115 distinct cited domains — and foursquare.com appears zero times. So do yelp.com, google.com and facebook.com. Here is what a citation count can and cannot prove, and the ranked list of domains that actually carry the answers.

First-party measurement, 2026-07-03 to 2026-07-31, read from the Search Console API: 25 of our URLs held an average position of 10 or better, together earning 308 impressions and exactly 1 click — a 0.32% click-through rate. A conservative page-one CTR curve predicts 12.6 clicks on that impression mix. The full 25-URL table, the arithmetic, and the two explanations we cannot yet separate.

Google's Search Console “Search generative AI” control is rolling out to a subset of website owners: one property-level switch that pulls your site out of AI Overviews, AI Mode and Discover's AI features without touching your rankings. It does not stop model training, it does not affect ChatGPT or Perplexity, and it is a different lever from both nosnippet and Google-Extended — three controls with three different scopes, read here from Google's own documentation. Plus a first-party check: of the 20 domains AI engines cite most in our corpus, only 4 of the 19 we could measure name Google-Extended at all.

geoSurge published “AI Searches What It Remembers” on 2026-07-28: across 66 buyer prompts run 60 times each, Gemini 3.5 Flash searched for brands already in its top-10 memory 3.2× as often as brands it didn't remember (55.7% vs 17.4%). That study is about national brands and the search step. We ran the equivalent test on the answer step for local SMBs — 163 audits, 53 metro+vertical groups, identical prompts, ChatGPT and Perplexity — and a business already named in another business's answers was named in its own at 27.7% vs 8.0%. A ratio of 3.5×.

On 2026-07-30 Google's Search Off the Record confirmed the old webmaster-guideline requirement to block internal search results pages is no longer in Search Essentials — it's now advice, not a rule. We then measured the AI-citation side on our own corpus: across 1,048 answers, 3,252 cited URLs and 1,140 domains, exactly 0 were internal search results pages, and only 1.11% carried any functional URL parameter at all. Blocking /search costs nothing in AI visibility we can measure.

We parsed the actual rules — not just the group headers — inside the robots.txt of the 20 most-cited domains in our audit corpus, on 2026-07-30. Of 70 (domain, agent) pairs where a group exists, 36 are Disallow: / with no exception, 27 are partial, and 7 explicitly Allow. Two domains were unmeasurable and are reported as such. And the blocking says almost nothing about whether the domain gets cited.

Victorious' Q2 2026 Quarterly Search Report (July 29, 2026) tested 175 brands across five verticals on eight AI platforms. 96% were described accurately when asked about by name; 89% never appeared in AI answers to category research questions. Our own 168-business dataset found the same shape: 98.8% recognised by name, 19.8% named in the buyer's question.

Pillarbase Research analysed 15,699,298 Google AI Mode citations across 148 industries. 47.7% carry a scroll-to-text highlight, and they resolve to 4,612,626 distinct passages across 2,738,150 pages. Median highlighted passage: 117 words, 80% answering in the first sentence, 83% self-contained. The unit AI Mode reuses is the paragraph, not the URL.

We swept all 99 domains that two or more of ChatGPT, Perplexity, Gemini and Claude cite on small-business AI-visibility questions, looking for a contact address. 7 publish an address belonging to a person, 68 publish only a role alias, 23 publish none at all. Method: 8 conventional contact paths per domain, raw-HTML extraction, own-domain addresses only, on 2026-07-30. The dataset and the classifier's rules are published with it.

On identical questions, Perplexity returned 546 citation entries to ChatGPT's 252 — more than twice as many. But Perplexity almost never types those links into the answer text: across 525 Perplexity answers in our production corpus, 11 carried an in-text URL, against 523 of 523 for ChatGPT. Any AI-visibility tool that reads the answer text instead of the citation payload is structurally blind to the engine that cites the most.

Google made Search Console platform properties globally available on 29 July 2026, reporting how your Instagram, TikTok, X and YouTube posts perform in Search, Discover and News. We checked the same four platforms against our own AI-citation corpora. Instagram, TikTok and X: zero citations in 144 four-engine answers and zero across 1,112 domains in 523 local ChatGPT answers. YouTube is the exception — and it is 21 of 36 Perplexity answers against 0 of 36 from ChatGPT and 0 of 36 from Claude.

We put 12 buyer questions through ChatGPT, Perplexity, Gemini and Claude, three samples each: 144 answers, 1,465 citation entries, 538 distinct domains. 439 of those domains (81.6%) were cited by exactly one of the four engines. Only 10 (1.9%) were cited by all four. Method is the engines' returned citation arrays, not a substring test on the answer text — and our own domain appears in zero of the 1,465.

Over 25-27 July 2026 shared Claude conversations turned up in Google. Anthropic's robots.txt had disallowed /share/* the whole time — we fetched it ourselves. Per Google's own documentation a robots.txt disallow never prevented listing, and it blocks the rule that does. The same mechanic runs the other way: in our 520 ChatGPT answers, Yelp blocks OAI-SearchBot and is cited 0 times, MapQuest blocks the same crawlers and is cited 12 times.

On July 14, 2026 Germany's ZAK issued its first orders against AI services, ruling that Google's AI Overviews and Perplexity are content providers, not neutral intermediaries, and that German media law applies. The orders demand transparency and non-discrimination in how links are made visible, where journalistic content sits relative to the AI answer, and how third-party sources are labelled. It is not yet binding and Google is appealing — but it is the first time any authority has tried to force an answer engine to state its selection criteria.

A July 2026 BrandMentions study of ~410,000 mentions found 50–60% of the third-party mentions around highly AI-visible brands are text-only with no outbound link. We measured the same thing on the output side. Across 520 ChatGPT answers to real buyer questions about 176 local businesses in 41 US metros, ChatGPT named 25 review and directory brands 639 times and linked them 481 times: 158 brand-answer pairs (24.7%) were a name with no link anywhere in the answer. The unlinked ones are the brands buyers have heard of — Yelp 57 named / 0 linked, Super Lawyers 26 / 3, Avvo 22 / 9. Full table and method.

Previsible's 2026 State of AI Discovery Report (July 6, 2026), built on 6.77 million LLM sessions across 166 GA4 properties, found that 28.8% of the traffic ChatGPT refers lands on a site's internal search results page rather than on a page that answers the question. Across all AI-referred traffic it's roughly 25%; in SaaS, 34.6%. Here's what that pattern means — the engine trusts your domain and can't pick a page — and why it's the same failure mode we measure on the recommendation side.

On July 21, 2026 OpenAI announced the ChatGPT for small business program: virtual training, in-person AI academies, new guides, and partner agents from Dropbox, Shopify, Intuit, Slack, Atlassian and Wix. Every component is about a small business using ChatGPT. None is about a small business being found in it — there is still no listing, no claim flow, no business profile. What that means for how ChatGPT decides which local business to name.

New first-party data (July 26, 2026): across 511 ChatGPT answers about 173 real local businesses in 42 US metros, 980 of the 1,092 cited domains (89.7%) appeared in exactly one metro. Only 112 domains crossed metros at all. We then checked 46 local-media domains — the city magazine, the daily paper and the alt-weekly for 15 metros — against the 379 answers from those same metros: 6 citations, in 4 of the 15 cities. Eleven metros got zero. Full per-metro table, the exact domain list checked, and the method.

Axios reported on July 23, 2026 that Yelp is licensing reviews, photos and business data to OpenAI for ChatGPT. Our pre-deal baseline, from 173 audits of real local businesses run June 12–July 14, 2026: across 511 ChatGPT answers citing 1,092 distinct domains, yelp.com appeared in exactly none of them — while ChatGPT mentioned Yelp by name in 11.0% of those answers without linking it once. Yelp's robots.txt, fetched July 25, blocks OAI-SearchBot from everything except /article. This deal is about access, not ranking, and it is not a surface anyone can optimise into.

New first-party data (July 25, 2026): across 511 ChatGPT web-search answers about 173 real local businesses in four verticals, the engine attached 5,081 links to 1,092 distinct domains — 9.9 links and 5.5 distinct sources per answer. 1,057 of those domains (96.8%) were cited in only one of the four verticals and just 12 appeared in all four. Four domains cover 53% of all answers, but it is a different four in each vertical: dentists lean on Birdeye and ThreeBestRated, med spas on MedSpaScout and AuraVenu, home services on Expertise.com and BBB, injury law on Birdeye and LawLeaderboard. Full ranked table and method.

OpenAI shipped Health in ChatGPT to all US adults on July 23, 2026 — connected Apple Health and medical records, drawn on across ordinary conversations. Its own post gives one local-recommendation example: choosing a restaurant around a dietary restriction. That is the first vendor-stated case of a local recommendation conditioned on data the business cannot see. Plus our baseline: across 333 ChatGPT answers about health-adjacent local businesses, only 15.6% cited any medical or provider-credential source.

Google added a new guideline to its review snippet structured-data docs on July 24, 2026: don't include fake or undisclosed incentivized reviews on your page or in your markup. The rule governs review content on your own site. In our corpus of 733 ChatGPT answers to local buyer questions, the engine cited the business's own domain in 29.5% of answers and a third-party review platform in 43.4% — and Yelp zero times in 3,787 cited domains.

New first-party data (July 24, 2026): 57 identical local buyer questions asked 3–8 times each over ~2 weeks. On ChatGPT only 4.2% of the businesses named survived every ask and 67.3% appeared exactly once; on Perplexity it was 15.4% and 50.9%. Mean pairwise overlap between two asks of the same question: 0.20 on ChatGPT, 0.42 on Perplexity. Volatility is not a property of 'AI' — it is a property of the engine you are reporting.

Google announced Gemini 3.6 Flash, 3.5 Flash-Lite and 3.5 Flash Cyber on July 21, 2026, and said plainly: "3.5 Flash-Lite is also rolling out in Google Search." It did not say AI Overviews or AI Mode. What a generator swap does change (who gets named, how sources are used), what it does not (the corpus the engine retrieves from), and the sampling-parameter deprecation that makes run-to-run variance harder to suppress for anyone measuring Gemini through the API.

Notified and PRWeek published "AI Is Your Audience" on July 22, 2026, reporting that 99.3% of press releases distributed via GlobeNewswire were cited by ChatGPT or Claude, with an average eight hours to first citation. We checked our own corpus: across 511 ChatGPT answers to local buyer questions, press-release wire domains were 6 of 2,815 source citations — 0.21% — with zero from GlobeNewswire or Business Wire. Both findings can be true. The gap between them is the thing to understand before spending a budget.

New first-party data (July 23, 2026): across 1,024 AI answers to local buyer questions, directory brands — Yelp, BBB, Zocdoc, Angi, Justia — were named as a recommendation 11 times out of 4,015 business mentions (0.27%). But directory domains are 25.4% of every source ChatGPT cited, they appear in 73.6% of its answers, and 12 of the 20 most-cited domains are directories. The listings are the input; the businesses are the output.

An open-access study (14,140 API runs, 12 brands, 5 AI engines, 2 category framings) found the brand with the strongest Knowledge Graph entity — 2.5x the next-highest — received 0% of recommendations in one category, while a brand 79x weaker received 93%. Change one word in the category framing and the recommended set changes. Own-brand websites were under 2% of all cited sources.

Semrush tracked 1,094 US categories in ChatGPT from January to June 2026 — 50,000+ brands, 600,000+ citations. Only 15.2% had a clear category owner; 53.7% were unsettled. In the highest-demand half of topics, which carries 98% of AI search volume, ownership fell to 11.3%. No classic SEO metric predicted ownership better than ~56%. Our own local-market data says the opposite is true one level down.

New first-party data (July 22, 2026): we put 293 identical buyer-intent questions to ChatGPT and Perplexity at the same time — 586 answers, 155 local businesses, 42 US metros. The two engines' recommended shortlists overlap by a mean of 10.1%; a third of question pairs share no business at all; and of the businesses named by any engine, 73.8% were named by exactly one. "AI visibility" measured on one engine is not a measure of AI visibility.

New first-party data (July 21, 2026): across 15 local markets where we audited 3+ competing businesses — 59 businesses, 343 ChatGPT and Perplexity answers — every market had a fixed incumbent shortlist. A mean of 4.7 rival names appeared in every audit of that market regardless of which business was the subject, capturing 34.6% of all rival mentions, while 62% of named rivals appeared in only one audit. And 58% of the businesses we audited were never named in a competitor's answers at all.

A critical survey of generative engine optimization (arXiv 2607.14035, July 15, 2026) reviewed 45 studies from Nov 2023 to July 2026 and concluded that “no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability.” General heuristics transfer poorly; only topical relevance and position replicate. What that means for anyone buying or selling AEO work in 2026.

A 90-day study by agency DeltaV Digital (Apr 14–Jul 13, 2026; 21,075 AI responses, 25,337 citations, 5 engines, 8 industries) found every industry has a distinct “citation fingerprint”: listicles are 61% of cited pages in B2B tech services, homepages 55% in local services, articles 54% in healthcare, program pages 53% in higher education. There is no single AEO playbook — the page type you must make citable depends on the vertical, which is exactly the problem for an agency serving many.

Two published citation datasets — 5W's State of AI Citations 2026 (680M+ citations across 12 studies) and Profound's platform-citation analysis — quantify how little AI engines overlap: only ~11% of domains are cited by both ChatGPT and Perplexity, only ~12% of AI-cited URLs sit in Google's top-10, and each engine leans on a different source type (ChatGPT/Wikipedia, Perplexity/Reddit). For an agency, that makes “get you seen in AI search” a per-engine claim, not one number — and it means a page-1 ranking barely predicts an AI citation.

SE Ranking analyzed 50,032 commercial keywords in Google AI Mode on June 30, 2026: 29.45% returned a text ad. The sharper finding for anyone selling AI visibility — only 11.53% of advertiser domains also appeared as cited sources, and just 2.32% of advertised URLs also ranked organically. Paid placement, citation, and rank are nearly disjoint surfaces. What that means for agencies pricing an AEO retainer.

Profound analyzed 32M+ google.com/searchviewer instances (Apr 15–Jun 30, 2026) and found google.com is now AI Mode's #2 most-cited domain, an 8.4x surge in citation share driven almost entirely by Google Business Profiles and Product Knowledge Panels. Hospitality, home services, restaurants, real estate and healthcare moved most. What it means for agencies whose clients are local service businesses.

CiteLens ran 320 buyer queries through ChatGPT, Perplexity, Claude and Google AI Mode and matched every cited source against Google's rankings. Google AI Mode drew 93% of citations from Google's top 10 (0.92 correlation) and Perplexity 89% (0.87). ChatGPT drew 30%, with near-zero correlation to rankings. One SEO playbook does not move all four engines — and here is the market caveat that keeps the number honest.

Semrush's 2026 AI Visibility Index analyzed 126 million US AI search prompts across ChatGPT, Gemini, AI Mode and AI Overviews. In News & Media the top three brands hold 82.9% of category visibility; in Finance, 41.4%. ChatGPT cites 15 sources per answer, Gemini cites 3. Only 36 brands were visible on all four platforms. What the concentration data means for agencies setting client expectations.

OpenAI launched Sites in ChatGPT on July 9, 2026: describe a site, ChatGPT builds and hosts it, you share it by URL. We read every page of the documentation on July 14 — publishing is documented, discoverability is not. Meanwhile chatgpt.com's own robots.txt blocks Google-Extended, PerplexityBot, and Claude by name. What that means before you host a client's content there.

A 9,946-run study published July 5, 2026 found ChatGPT routes searches through four undocumented retrieval sources. In 11.6% of prompts the source changed between identical runs — and when it did, URL overlap fell about 45%. Same question, different cited sites. Here is what that does to anyone measuring AI visibility.

Cloudflare announced Precursor on July 13, 2026: session-long behavioral detection of automated and agentic traffic. It lands ten weeks before Cloudflare's new September 15 defaults, which block Agent and Training crawlers on ad-serving pages for new domains. Your client's WAF is now an AI-visibility setting.

OpenAI has confirmed it is deprecating ChatGPT Atlas, its standalone browser, which stops working August 9, 2026. Browser-based agentic capabilities move into the ChatGPT desktop app and a Chrome extension. The agent that reads your website stops being a niche app and becomes a default feature.

OpenAI's ChatGPT Ads shipped an Ads Manager overview tab, custom audiences, suggested ad drafts and a new e-commerce card in July 2026, and is now live in Japan and South Korea. But OpenAI's own published rule has not changed: ads do not influence the answers ChatGPT gives. The recommendation slot is not for sale.

On July 6, 2026, YouTube made Ask YouTube — its conversational, AI-generated search experience — available to all signed-in US desktop viewers 13+, out of a Premium-only test. It answers comparison and recommendation questions with synthesized text plus video clips. It is a fifth answer engine deciding who gets named, and the only thing it can name you from is video.

A randomized field experiment by Saharsh Agarwal (ISB) and Ananya Sen (Carnegie Mellon), running 1,065 Chrome users and 68,089 searches, found Google's AI Overviews cut outbound organic clicks 39.8% and raised zero-click searches 34.5% — with no improvement in user satisfaction, and no evidence the vanished clicks were the low-engagement ones.

On July 9, 2026, OpenAI made its GPT-5.6 family (Sol, Terra, Luna) generally available across ChatGPT, Codex, and the API after a two-week government cybersecurity review. ChatGPT is the surface that drives the vast majority of trackable AI referral traffic — so any AI-visibility measurement of ChatGPT taken before July 9 now describes a model your buyers may no longer be talking to.

Semrush's expanded 2026 AI Visibility Index (June 26, 2026) analyzed 126 million US AI search prompts across ChatGPT, Gemini, and Google's AI surfaces. Two findings matter most: being mentioned by an AI is not the same as your site being cited (on Gemini the overlap is as low as 30%), and 45% of marketing leaders can't accurately measure their AI visibility while only 9% have the tools to.

On July 1, 2026, Cloudflare launched its Attribution Business Insights dashboard, exposing crawl-to-referral ratios for AI bots that run from 118:1 up to nearly 50,000:1 — thousands of pages read for every visitor sent back. For local businesses it confirms the hard truth: the click isn't coming, so being the recommended name is the entire return.

Previsible's 2026 State of AI Discovery Report (July 6, 2026), built on 6.77 million LLM sessions across 166 GA4 properties, found ChatGPT drives 92.4% of standalone-LLM referral traffic — with Claude up 64x and Google's AI Overviews excluded entirely. Referral clicks and being recommended are different channels, and a GA4 traffic report measures neither well.

Google Search Console's new Generative AI performance reports (launched June 3, 2026, expanding globally through July) finally show how often your links appear in AI Overviews and AI Mode. But they report impressions only — no clicks, no queries, and nothing about the three engines outside Google or whether the model actually recommended you.

A Q2 2026 study of 1,008 consumers by Fractl and Search Engine Land found AI-search usage still rising while the share who call it 'more helpful' than traditional search fell from 82% to 54% in a year, and the share who say heavy AI use lowers their trust in a brand nearly doubled to 39%. Being named by AI is no longer enough — being accurate is.

Google's AI Mode passed 1 billion monthly users and is rolling out agentic calling and booking to everyone in the US this summer 2026. When the AI acts on one business's behalf, ranking on page two no longer matters — being the single name it chose does.

Every Monday we run our own audit engine against 12 buyer questions across ChatGPT, Perplexity, Gemini and Claude. The 2026-09-07 run: mention rate 2.1% (3 of 144), citation rate 3.5% (5 of 144), share of voice 0.36%, down from 5.6% last week. The pre-registered reads: Perplexity attached our Profound-alternatives page 3 of 3 for the fifth week (15 of 15 samples); the two comparison pages ChatGPT cited for the first time on 08-31 went 0 of 3 each; ChatGPT fell from 5 to 2 of 36. Gemini and Claude still 0. Per-engine numbers every week, including the down weeks.

Anthropic made Claude Sonnet 5 the default model for Free and Pro users on June 30, 2026. Claude's answers about businesses now come from a new model — visibility numbers measured before the swap are stale.