Week eighteen: 5 citations in 144 answers, against 7 last week and 7 the week before, on the same count. ChatGPT fell from 3 to 2; Perplexity fell from 4 to 3, every one the Profound-alternatives page. Gemini and Claude 0 for the eighteenth week

The panel and the method

The self-audit is deliberately boring and fixed, so the number is comparable week to week. Twelve pinned buyer questions — “What are the best AI visibility tools for small businesses?”, “Is there a one-time AI visibility audit instead of a monthly subscription?”, “What is a good Profound alternative?” and nine more — run three times each against four engines (OpenAI, Perplexity, Gemini, Anthropic). That's 144 answers per run. We record whether AskedAbout is mentioned, whether it's cited with a link, and share-of-voice. The engine is the exact production engine that powers a paying customer's audit, so the number we report on ourselves is the number a customer would see on themselves.

The numbers (runs of 2026-10-19, 2026-10-12 and 2026-10-05, all on the new count)

The Search Console row is the whole askedabout.com property, not this post's URL, which has shown zero web impressions in every window since 2026-07-01. The stored mention series since tracking began 2026-06-27 is 0, 0, 0, 0, 1, 2, 0, 0, 4, 5, 5, 3, 3, 3, 7, 5, 3, 2 across eighteen scheduled weekly runs (plus two off-schedule zero runs on 2026-07-28, disclosed below). The old-count citation series ended at 0, 0, 0, 0, 1, 2, 0, 3, 7, 8, 8, 5, 6; the new count reads 3, 6, 7, 7, 5. We are not joining the two into one line. Nothing in the engine's run or its citation parser changed between the five new-count runs. Against 10-12 the stored URL totals moved 1.6% on ChatGPT (249 to 253), 5.9% on Perplexity (220 to 207), 4.6% on Gemini (479 to 501) and 17.6% on Claude (136 to 112). The Claude move is above the 15% band we watch; it is the engine's own reference count, not ours, and the Claude citation count for us is 0 on both sides of it.

2 stored mentions, 2 of them ours; 5 citations, 3 of them with no mention of us in the text

The engine's judge reads each answer and records whether it mentions AskedAbout. This week it recorded 2. As a check that does not depend on the judge, we search every answer's text for our name or domain: 2 by string. On 09-28 and 10-05 the two counts differed by one each week, and each time the odd row linked a look-alike domain (askedaudits.com, then askedbyai.com). On 10-12 the judge stored 3 and the string count was 3, the two equal. This week there is no judged row without our string, so the stored count stands as the mention count.

The table lists every answer in this week's run that the judge counted, that carries our string, or that stores one of our pages as cited. The 3 Perplexity rows on the Profound question cite the page as a numbered source and never name AskedAbout in the text: citations, not mentions, and the judge stored them as unmentioned.

Three counts, then, on one run: stored by the judge 2, our string in the text 2, the product named as a tool 2, and 5 answers citing a page of ours. For anyone reading a mention count from any tool, ours included: ask whether a mention is decided by a model or by a match, whether a citation requires the name in the text, and ask to see the sentence. On a base of 144 answers the counts above span 1.4% to 3.5%.

The pre-registered reads, two weeks of them

The 10-05 update pre-registered five reads for 10-12; the page was not updated that week, so both weeks are read here. ChatGPT's count from 1 of 36: 3 on 10-12, 2 this week, both inside the 1 to 6 band of the eight readings before this one. By question over the last three runs: "How can I check what ChatGPT says about my business?" 1 → 2 → 1 of 3; "What tools track whether AI recommends my business?" 0 → 0 → 0; "Which tools show how ChatGPT, Perplexity and Gemini recommend businesses?" 0 → 0 → 1; "What is a good Profound alternative for a small business?" 0 → 1 → 0; "How do I improve my business's visibility in ChatGPT and AI search?" 0 → 0 → 0. The homepage was cited in 2 answers this week, 2 on 10-12, 1 on 10-05. Perplexity's count from 6 of 36 and whether /compare/best-profound-alternatives holds on both questions: 4 on 10-12 and 3 this week. On the Profound question the page held, 3 of 3 and 3 of 3; on the one-time-audit question it did not, 3 of 3 on 10-05, 0 of 3 on 10-12, 0 of 3 this week, so the page's second question lasted one run. That question did cite us once on 10-12, through /compare/cheapest-ai-visibility-tools (1 of 3), one of the two comparison pages this series had read as lapsed on their own ChatGPT questions since 09-07; this week 0 of 3. Whether the judge and the string match agree: 10-12 yes (3 against 3), this week yes (2 against 2). The sixth and seventh reads on the two lapsed comparison pages, on their own ChatGPT questions: /compare/cheapest-ai-visibility-tools 0 → 0 → 0 and /compare/athenahq-alternative 0 → 0 → 0 of 3, seven weeks at 0 for both on those questions; the cheapest page's one Perplexity citation on 10-12 came on a different question and did not repeat, which is one sample against our lost-citations study and no more than that. Perplexity's stored URL total against 204: 220 on 10-12, 207 this week.

The pairs that moved this week, against 10-12: dropped: ChatGPT on "How can I check what ChatGPT says about my business?" 2 → 1 of 3; Perplexity on "Is there a one-time AI visibility audit instead of a monthly subscription?" 1 → 0 of 3; ChatGPT on "What is a good Profound alternative for a small business?" 1 → 0 of 3; new: ChatGPT on "Which tools show how ChatGPT, Perplexity and Gemini recommend businesses?" 0 → 1 of 3. Perplexity's three answers on the Profound question referenced 5, 5, 6 sources this week (7, 6, 7 on 10-12), ours among them each time, and none of the six names us. On the one-time-audit question this week's three answers referenced 5, 5, 5 sources, none of ours, and named AskedAbout in 0 of 3 (10-12: 7, 8, 5 sources, 1 of 3 citing a page of ours, 0 of 3 naming us). Perplexity's five new-count readings are 0, 0, 6, 4, 3.

What changed in the count on 2026-09-20

An engine's API returns two things with an answer: the text, and a list of sources. Perplexity's list held 17 to 20 URLs on every answer we stored; the text references some of them with numbered markers like [2] and leaves the rest unreferenced. Until 2026-09-20 our engine stored the whole list as the answer's citations. Since then (commit 4ceb08f) it stores as cited only what the text references: a numbered marker on Perplexity, an annotation or link on the others. Sources that were returned and not referenced are retrieved, not cited. New customer audits use the same rule from the same date. The same release replaced the judge that reads each answer: it now reads the whole answer, where the old one read the first 6,000 characters, and it lists only brands in our category.

The size of the change, from the two runs either side of it: Perplexity went from 720 stored URLs on 09-14 (20 on each of 36 answers) to 285 on 09-21 (between 1 and 15 per answer, median 7.5); Claude from 222 to 127; Gemini from 498 to 431; ChatGPT from 265 to 259. As a cross-check we applied the marker rule to the stored 09-14 Perplexity answers: 321 of the 720 listed URLs (44.6%) are referenced by a marker in the text, against 285 stored as cited this week. The weekly cited-domain series on the same answers has the same break; that post carries its own update.

The six-week Perplexity streak, re-read with the new rule: 7 of 18

From 08-10 to 09-14 this page reported that Perplexity attached askedabout.com/compare/best-profound-alternatives in 3 of 3 samples on "What is a good Profound alternative for a small business?", 18 of 18 samples over six weeks, and every week noted that the answer text never named us (mentioned: false on every row). The answer text and the source list for all 18 samples are stored, so the new rule can be applied to them after the fact: find the page's position in the list, then look for that number as a marker in the text.

The page sat at position 2 to 5 of the list in all 18 samples. The text referenced it in 7: all three samples on 08-24 and 09-14, one on 08-31, none on 08-10, 08-17 or 09-07. So what we called an 18-of-18 citation was an 18-of-18 retrieval and a 7-of-18 citation, and the weekly reading under the new rule would have been 0, 0, 3, 1, 0, 3 of 3, not a flat 3. The 0 of 3 on 09-21 and 09-28 had precedent in that series on three of six weeks, and on 2026-10-05, 2026-10-12 and 2026-10-19 the text referenced the page in 3 of 3, 3 of 3 and 3 of 3 samples. One limit: the weekly run now stores only what the text references, so we cannot say whether Perplexity still retrieves the page.

For anyone reading a Perplexity API response, or a tool built on one: the source list returned with an answer is longer than the set of sources the answer text references with numbered markers. Across the seven stored runs, 2,021 of 4,979 listed URLs (40.6%) were referenced by a marker. A count of listed URLs overstates in-text citations by roughly that ratio, and ours did for six weeks.

The honest caveats, before next week's read

First: 2 answers in 144 carry our name or domain, 2 name the product, 5 cite a page of ours, and 139 do none of those; Gemini and Claude have never once named or linked us in eighteen runs. Second: 7 to 7 to 5 is three totals that hide the moves inside them: ChatGPT 1 → 3 → 2, Perplexity 6 → 4 → 3, and the same question can cite us in 3 of 3 samples one week and 0 of 3 the next, as the streak table above and the one-time-audit question this week show. Third: our judge miscounted one row on 09-28 and one on 10-05, through two look-alike domains; a second, simpler count caught both, and on 10-12 and this week the two counts agree. Fourth: the engines' own reference counts move from week to week with no change on our side, so every per-engine reading sits on a different base of references per answer. Fifth: API samples are samples. They are not proof of what every person asking the same question in the app sees. Pre-registered for 2026-10-26: ChatGPT's count from 2 of 36 (the last seven weekly readings on ChatGPT were 2, 3, 3, 6, 1, 3, 2); Perplexity's count from 3 of 36 and whether /compare/best-profound-alternatives holds on the Profound question for a fourth week; whether the judge and the string match agree; the eighth read on the two lapsed comparison pages; and the four stored URL totals against 253 / 207 / 501 / 112, as a check on the engines. All of it goes in this table whichever way it lands.

One sample from a distribution

That is the finding, and it is worth stating plainly because we sell measurement: a single AI-visibility check — including a single check of ours, including a week whose citation count we had to restate — is one sample from a distribution. In our churn study across 57 repeated buyer questions, the mean overlap between two asks of the identical question on ChatGPT was 0.20 — four-fifths of the shortlist changes between asks, and only 4.2% of named businesses survived every repeat. Our own 1 → 2 → 0 → 0 → 4 → 5 → 5 → 3 → 3 → 3 → 7 → 5 → 3 → 2 is that distribution, experienced from the inside — a flat topline over a churning composition. Any tool or agency that showed you a one-off screenshot of either our good week or our bad week would be telling you a true thing that is also meaningless. The panel, repeated on schedule, is the measurement.

Who the engines name instead

The more useful half of the data is what the engines do say when asked these questions. This week's most-named brands, counted per answer across the run: otterly.ai (51), peec ai (40), profound (35), athenahq (22), rankscale (20), chatgpt (17), ahrefs brand radar (15), llm pulse (15), semrush ai visibility toolkit (14), hubspot aeo (14). On 10-12, under the same judge, the order was otterly.ai (44), peec ai (37), profound (33), semrush ai visibility toolkit (27), ahrefs brand radar (26), athenahq (23), chatgpt (19), otterly ai (17), perplexity (15), brightlocal (14). The specialists week over week: Otterly.ai 44 → 51, Peec AI 37 → 40, Profound 33 → 35, AthenaHQ 23 → 22. Entered the top ten: rankscale, llm pulse, hubspot aeo. Left it: otterly ai, perplexity, brightlocal. No specialist clears a third of answers. The tally comes from the same judge that miscounted one row of ours on 09-28 and on 10-05, so read each count as approximate by a row or two. On the one question in this panel that asks it directly, our four-week ranked list of the tools the engines name has Otterly.ai first, with AskedAbout cited on it 0 of 3 this week.

One instrument fix, disclosed

On 2026-07-28 we found and fixed a bug in our own engine's citation sensor: it was recording source URLs from only one of the four engines, so earlier weekly rows understated how many sources the engines actually consult. Post-fix, the runs have recorded 1,587 → 1,579 → 1,637 → 1,652 → 1,618 → 1,652 → 1,705 → 1,705 source URLs per run through 09-14 (Perplexity alone: around 700–710 every week). The fix does not touch the mention numbers above — whether an answer names AskedAbout was measured the same way all eighteen weeks — but the two hand-triggered 07-28 runs in our series exist because we re-ran the audit to validate the repaired sensor rather than wait six days for the cron. Both read zero mentions. We say this because a measurement company that silently repairs its instruments deserves the skepticism it would earn.

The 2026-09-20 change above is the second instrument change in this series. It moves every citation number and the brand tally on this page; the mention count is cross-checked by the string match above. The old method still exists in the code as an explicit option, so a past baseline can be repeated on its own terms; the weekly run uses the new one.

What we changed in the two weeks, and the takeaway

Between the 10-05 and 2026-10-19 runs the instrument did not change. The site did. The page registry that holds every /compare and /guide page had 3 commits (2ac8028, 42bee07, 22e24ff) in the span, each a repair to one existing page by the growth role. The Profound-alternatives entry was edited once in that span, on 10-06: the Rankscale and Rankability price rows re-read from the vendors' pages on 10-01 (Rankscale $20 to $17 a month, yearly only; Rankability's public price withdrawn), nothing else on the page; the how-to-check guide was not edited. /news published 3 new posts in the two weeks (2026-10-12, 2026-10-14, 2026-10-15). We cannot credit any of it with the Perplexity page's third week on the Profound question, or blame any of it for the one-time-audit question dropping the page after one run: the one edit was two price rows on 10-06, six days before a run on which the page's first question held at 3 of 3 and its second dropped it, and the two questions moved in opposite directions on the same page.

The takeaway for any business reading this: a total can hold still while every part of it moves, and it can move while the page under it does not. Our citation count went 7 to 7 to 5; underneath it Perplexity kept one page on one question for three weeks and dropped the same page on another question after one, and ChatGPT swapped which of our questions it cites from week to week. A tool that reports the total would call the last two weeks a slide. A tool that reports the engine, the page and the question would call it the same page holding on one question and churning on the others. We report both, in public, on schedule, including the weeks our own judge needs a correction. You can measure yours in 60 seconds, and if you're evaluating tools to track it over time, here's how the category compares, including where we sit: cited by two engines on 3 of twelve questions, named on 2, absent from the other two engines, and back on the panel next Monday.