AI search news ·
ChatGPT and Perplexity answer the same buyer question with almost entirely different businesses
Every audit we run puts the identical buyer-intent question to more than one engine at the same moment — same wording, same city, same category, same day. That makes a comparison possible that a single-engine tool cannot make: not how often each engine names a business, but whether the two engines are even talking about the same set of businesses. Across 293 identical question pairs — 586 answers, 155 local businesses, 42 US metros, 12 June to 11 July 2026 — the answer is no. Mean overlap between the two recommended shortlists is 10.1%; 33.1% of question pairs share zero businesses; and of the 65 businesses named by any engine on a buyer question, 48 (73.8%) were named by exactly one of the two. ChatGPT names a mean of 6.2 businesses per answer, Perplexity 5.2, and they agree on about one.
Primary source: AskedAbout (first-party audit corpus, 173 audits)
What we measured, and how
Our production corpus is 173 completed audits (synthetic QA rows removed by an exact-name filter, not a regex over the vertical field — a discipline adopted after we published a false finding in July and corrected it publicly). Each audit poses buyer-intent questions — "What are the best med spa in Dallas, TX?", "I'm looking for a personal injury law firm in Houston, TX. Who do you recommend and why?" — to ChatGPT and Perplexity, and records the businesses each engine names in each answer. For this cut we kept only question pairs where both engines answered the same prompt in the same audit, and dropped the branded control questions ("is [Business] a good choice?"), which test recall rather than recommendation. That leaves 293 pairs / 586 answers spanning med spas, cosmetic and family dentists, personal-injury law, plumbers, HVAC, roofing and electricians in 42 US metros, run between 2026-06-12 and 2026-07-11. The question this cut asks is the one a single-engine check structurally cannot: when you change the engine and change nothing else, does the recommended set change?
Finding 1: the two shortlists overlap by about one name in ten
For each pair we compared the sets of businesses named, matching names leniently — normalised case and punctuation, legal suffixes (LLC, LLP, P.A., P.C.) dropped, and a match counted when one name's tokens are a subset of the other's or the two share 60%+ of their tokens. That means "Zehl & Associates" and "Zehl & Associates Injury & Accident Lawyers" count as agreement, not disagreement. The lenient reading is the one reported below, because it is the one that flatters the engines:
| Measure (293 identical question pairs) | Lenient match | Exact-name match |
|---|---|---|
| Mean set overlap (Jaccard) | 10.1% | 6.2% |
| Median set overlap | 9.1% | 6.7% |
| Pairs sharing zero businesses | 33.1% (97) | 49.5% (145) |
| Mean businesses named — ChatGPT | 6.2 | 6.2 |
| Mean businesses named — Perplexity | 5.2 | 5.2 |
| Mean businesses both engines named | 0.97 | 0.60 |
Roughly eleven distinct businesses get named across the two answers, and about one of them is named by both. This is not a ranking difference — it is a different cast of characters. It is also consistent with what published third-party citation research finds one layer down, at the level of sources rather than businesses: only ~11% of cited domains are shared between ChatGPT and Perplexity. Different corpora, different unit of analysis, same shape of answer.
Finding 2: 74% of the businesses with any AI visibility have it on one engine only
Set overlap is a property of the answer. The commercially relevant question is a property of the business: if an engine names you, does the other one? Of the 155 businesses in this cut, 90 (58.1%) were named by neither engine on any buyer-intent question — which independently reproduces the 58% we measured on a different cut of the corpus in AI knows your business, it just never brings it up. Of the 65 that were named by at least one engine:
| Visibility pattern | Businesses | Share of the 65 visible |
|---|---|---|
| Named by both engines | 17 | 26.2% |
| Named by ChatGPT only | 23 | 35.4% |
| Named by Perplexity only | 25 | 38.5% |
| Single-engine visibility (either) | 48 | 73.8% |
Neither engine is the generous one — ChatGPT-only and Perplexity-only are within 3 points of each other. The asymmetry is not between the engines; it is between one engine and two. At the answer level the same thing shows up sharper still: of the 97 question pairs where either engine named the subject business, 76.3% were single-engine.
Finding 3: what it looks like in one market
A worked example, so the number is not an abstraction. The question — "What are the best personal injury law firm in Houston, TX? Give a short ranked list with a one-line reason each." — asked of both engines, in the same audit, on the same day:
- ChatGPT: Arnold & Itkin LLP · The Lanier Law Firm · The Ammons Law Firm · The Buzbee Law Firm · Williams Hart & Boundas · The Krist Law Firm · Zehl & Associates
- Perplexity: Perdue & Kidd · Zehl & Associates · Sutliff & Stout · The Hadi Law Firm · Abraham Watkins · Byron Alfred & Associates · Ben Crump Law · Doan Law Firm
Fifteen firm-slots, fourteen distinct firms, one — Zehl & Associates — on both lists. Six of the seven firms a Houston buyer would see in ChatGPT do not exist in Perplexity's answer, and seven of eight in Perplexity's do not exist in ChatGPT's. There were 85 such near-disjoint pairs (both engines naming 5+ businesses, zero exact-name agreement) in this cut alone.
What this means if you sell AI visibility to clients
- A one-engine report is not a baseline. If you check a client in ChatGPT and report "you're invisible in AI," you have a 35% chance of being wrong in the direction that costs you credibility later, and a 38% chance of missing the engine where they already win. Report engines as separate numbers, never averaged into one "AI visibility score."
- Absence on one engine is a finding, not a failure. Single-engine visibility is the norm (73.8%), not the exception. A client who appears in Perplexity and not ChatGPT is in the majority — the honest framing is a coverage gap with a named target, not a diagnosis of doom.
- Scope the work per engine. The two engines assemble answers from largely different source sets, so the fix for a Perplexity gap (review aggregators, directory presence, Reddit-adjacent surfaces) is not the fix for a ChatGPT gap. Optimising for "AI" as one thing is optimising for an average that exists nowhere.
- Re-measure, don't re-check. A single answer is one draw from a distribution — only 31.4% of recommended brands survived every repeat of the identical question. Two engines multiply that variance rather than cancelling it.
Limits — what this cut does not show
- Two engines, not four. ChatGPT and Perplexity only. Gemini and Claude are in the paid audit but not in this cut; our earlier four-engine study found Gemini and Claude name businesses at roughly twice ChatGPT's rate, so a four-engine version of this analysis would likely show more fragmentation, not less.
- One sample per question per engine. These are single draws, not repeated samples, so part of the measured disagreement is run-to-run variance rather than a stable engine difference. That is a reason to distrust any single check — including a single check on both engines — and it is why the paid audit samples each question three times.
- Name matching is imperfect. Businesses are matched by name string, so a rebrand or a heavily abbreviated name can read as two businesses. The lenient matcher above is our correction for that, and both readings are published so you can see the size of the effect (10.1% vs 6.2%).
- Convenience sample. These are businesses we chose to audit, not a random sample of US local businesses, and the vertical mix is weighted toward med spas and personal-injury law. Directionally strong, not nationally representative.
- We sell this measurement. AskedAbout sells AI-visibility audits, so read the numbers and discount the conclusions — ours included. The method above is stated in full precisely so it can be checked against your own clients.
You can run the free 60-second check on any business and see its mention rate on ChatGPT and Perplexity as two separate numbers — which, on this data, is the only honest way to report it. Agencies baselining a book of clients can run five at once with the $249 Agency 5-pack: 25 buyer questions × 4 engines × 3 samples per business, white-labeled, which is the repeated-sample version of the measurement this post is a single-sample cut of.
Does this mean one of the engines is wrong?
No — and that framing is the trap. There is no ground-truth ranking of "the best med spa in Dallas" for an engine to be wrong about. Each engine assembles its answer from a different set of sources it trusts, so the two are answering the same question from different evidence. The practical consequence is not that one is wrong; it is that a business's presence in one tells you very little about its presence in the other.
Which engine should a local business optimise for?
Whichever one its buyers use, measured rather than assumed — and then the other one, separately. In this cut ChatGPT-only (35.4%) and Perplexity-only (38.5%) visibility are nearly equally common, so there is no engine that reliably serves as a proxy for the rest. Start by measuring both as separate numbers.
Is 293 question pairs enough to conclude this?
For the headline effect, yes — a 10% mean overlap with a third of pairs at literally zero shared names is not a subtle difference that a larger sample would reverse. For the per-vertical and per-metro breakdowns it is not; those cells fall to single digits fast, which is why this post reports the pooled number and publishes its limits rather than slicing until something looks dramatic.
How can I check this for my own clients?
Run the same business through both engines with the identical buyer-intent question — not the business's name — and compare the sets of businesses named, not just whether your client appears. If you want it done at scale, that is what our audit does: 25 buyer questions per business, four engines, three samples each, with the per-engine numbers kept separate.
See your number
A free 60-second check shows what AI says about you.
Running this for clients? The $249 agency 5-pack audits five businesses, white-labeled.
Who runs this
- Built and operated by Sensara LLC, Atlanta, Georgia — about us and how the audit works.
- See what the report looks like before you run anything — score per engine, the competitors AI names instead of you, and a fix plan.
- We run the same audit on ourselves every week and publish the result: in the latest run AI named AskedAbout in 1 of 144 answers. We report our own numbers the way we report yours.