We built a measurement harness that asks three AI engines the same 53 questions every day and records every source each one cites. Over 17 days they cited 1,378 different domains between them. Twenty-six of those were cited by all three. If you are told your business now shows up in AI, the honest follow-up question is: in which one?
of the 1,378 domains cited across the study were cited by all three AI engines · 26 domains in total
of cited domains were cited by exactly ONE engine and never by the other two
domain overlap between OpenAI and Perplexity · Profound measured 11.0% on a much larger sample
of questions where both OpenAI and Perplexity cited sources, they shared not one single domain
cited sources per answer · Perplexity against OpenAI. Perplexity cited on 100% of answers, OpenAI on 34%
engines cited Mode’s own site. The other two never did, in 17 days. Our miss, published
Source: Mode Marketing original research, 13 to 29 July 2026. 53 questions, 7 categories, replayed daily across OpenAI GPT-5, Anthropic Claude Sonnet 5 and Perplexity Sonar via each provider's official API. Full method below.
The industry talks about “AI visibility” as though it were one thing you either have or do not have, in the way a business either ranks on page one of Google or does not. Our own measurement says that framing is wrong, and not by a little.
Across 17 days of identical questions, the three engines produced a combined pool of 1,378 cited domains. Only 26 of those, 1.9%, were cited by all three. The overwhelming majority, 84.6%, were cited by exactly one engine and never once by the other two. Every engine is, in effect, reading a different internet.
The pairwise picture is where it gets uncomfortable for anyone selling a single AI-visibility number.
Read the last column slowly. On 61% of the questions where both OpenAI and Claude cited sources, they did not share a single one. On 32% of the questions, OpenAI and Perplexity shared none. These are not answers that differ in emphasis. They are answers built on entirely separate evidence.
A figure has been circulating hard this year: only 11% of domains are cited by both ChatGPT and Perplexity. We went looking for where it actually comes from before repeating it.
It traces to Profound, who ran 100,000 identical prompts through both engines and reported 11.0% shared domains, 37.4% cited only by ChatGPT and 51.6% only by Perplexity. That is a real study with a stated sample and a stated method, and it is the primary source. A second analysis, from Whitehat SEO, examined 118,000 answers across four platforms between January and March 2026 and also arrived at 11%.
Two independent studies landing on the same number looks like strong corroboration. It is worth noticing that they are not measuring quite the same thing. Profound's 11.0% is a two-engine pairwise overlap. Whitehat's 11% is the share of cited domains “appearing across multiple platforms” among four engines, which is a different denominator. Both are useful. They are not the same statistic wearing the same number by coincidence so much as two nearby measurements that have since been flattened into one industry fact.
Our own replication, on a much smaller and deliberately UK-specific question set, found 4.4% overlap between OpenAI and Perplexity. The direction replicates decisively. The magnitude is starker on our sample than on Profound's. We would not claim our 4.4% is the truer number, because 100,000 prompts beats 53 questions on statistical weight every time. What we would claim is that the honest reading of all three measurements together is “somewhere in the range of roughly 4% to 11%, and low wherever you look”, rather than a constant to be quoted to one decimal place.
The overlap numbers tell you the engines disagree. The source mix tells you why. Given identical questions, each engine reached for a visibly different class of evidence.
Government and official data 6.6% · trade bodies and company registers prominent · Wikipedia 0% · Reddit 0% · YouTube 0%
LinkedIn 9.9% · Wikipedia 5.9% · Reddit 0% · YouTube 0% · government and official 0.4%
YouTube 7.1% · LinkedIn 6.9% · Reddit 4.0% · Wikipedia 0.3% · government and official 1.4%
Two zeros are worth pausing on. Claude cited Reddit and YouTube not once in 977 citations. OpenAI cited Wikipedia and Reddit not once in 257. These are not small preferences. They are whole categories of the web that one engine treats as evidence and another does not acknowledge at all.
This is also where we part company with a claim doing the rounds, that ChatGPT is overwhelmingly Wikipedia-driven and Perplexity overwhelmingly Reddit-driven, at shares near 47%. We could not reproduce anything like those magnitudes: Wikipedia was 0% of OpenAI's citations here and Reddit was 4.0% of Perplexity's. The likeliest explanation is a denominator difference, since a share of an engine's ten most-cited domains is a very different figure from a share of all its citations. Our numbers are shares of all citations, stated plainly so they can be checked. The wider point holds regardless: the engines lean in genuinely different directions.
The question set includes questions about Mode, because we run this harness on ourselves first. Here is what 17 days returned.
OpenAI cited modemarketing.ai and modemarketing.co.uk. Perplexity cited neither, on any question, on any day. Claude cited neither, on any question, on any day. By the tidy version of this story we would be “visible in AI”. By the honest version we are visible in one engine out of three, and the other two would tell a prospective client we do not exist.
We publish that for the same reason we publish the rest: a company selling AI visibility that only ever shows you its wins is showing you marketing, not measurement. It is also the cleanest possible demonstration of the finding. If a business that builds measurement harnesses for a living is cited by one engine and invisible on two, the odds that your single flattering screenshot represents the whole picture are not good.
It tells us something useful, too. The engine that cites us is the one leaning on official records and published data, which is exactly the kind of source Mode has spent this year producing. The two that do not are the ones weighting community platforms and third-party mentions, which is the work we have done least of. The measurement points straight at the gap.
One AI answer is not a report card. Asking ChatGPT about your business and liking the answer tells you about ChatGPT. On this evidence it predicts very little about what Perplexity or Claude will say. Any check worth paying for has to run per engine, and any tool or agency reporting a single AI-visibility score owes you an explanation of which engines are inside it.
Some doors are wider than others. Perplexity cited 15 sources on every single answer. OpenAI cited 5, on a third of answers. That is not a subtlety, it is a difference in how many slots exist to be won. If you are starting from nothing, the engines that cite generously are where movement shows up first, and the selective ones are where it shows up last.
Chase the signals that travel. The temptation, reading this, is to optimise per engine. We would not. Engine-specific tactics age badly, and the retrieval systems behind these answers are rewritten constantly. What sits underneath all three is duller and more durable: a site machines can actually read, unambiguous structured data about who you are, and real third-party mentions across the different kinds of source these engines favour. That is the same advice we would have given before running this study. The study is why we can now say it with numbers attached.
13 to 29 July 2026, 17 consecutive days. A fixed set of 53 questions across seven categories (brand, cost, trust, AI visibility, market, verticals, commercial intent) was replayed on a daily schedule to three engines through each provider's official API: OpenAI GPT-5 with web search, Anthropic Claude Sonnet 5 with web search, and Perplexity Sonar. Every engine received identical question text. For each response we recorded the list of source URLs the API returned, normalised those to domains (stripping www), and deduplicated per engine per question across the whole window. Overlap is reported two ways: as a share of the combined domain pool for a pair of engines, and as a per-question mean Jaccard similarity. No scraping and no bot evasion at any point · official APIs only.
Limitations, stated up front. API answers are not the same surface as the consumer apps most people use, and published work suggests the two overlap only partially, so read this as the API-surface picture rather than what a given person sees in the ChatGPT app. The question set is 53 questions weighted toward UK marketing, agency selection and AI visibility, which is our market and not a general-purpose sample · a different subject area could easily produce different source mixes. Sample sizes differ by engine, because OpenAI and Claude returned citations on only some answers (34% and 57%) while Perplexity returned them on all, so pairwise comparisons rest on 28 to 41 questions rather than all 53. A fourth engine, Google Gemini, was configured and called throughout but is excluded entirely: the API project's quota is unprovisioned, every call errored, and no citation data was collected. 133 of 708 calls errored across the study, mostly Gemini.
Our interest, declared. Mode sells AI visibility services, and a finding of “you have to check every engine separately” is convenient for a company that sells per-engine checks. Treat the interpretation with that in mind and the numbers on their own merits. The underlying dataset, including the full domain lists, is available on request at hello@modemarketing.ai.
Position: AI visibility is per engine, not a single status. Citation overlap between the major engines is low, and being named by one is weak evidence that you are named by the others.
Confidence: strong on the direction, moderate on the magnitude. Three independent measurements at very different scales (100,000 prompts, 118,000 answers, and our 53 questions) all land on low overlap, which is about as much agreement as this young field offers. The exact percentage is far less settled, and our own figure sits well below the widely quoted one.
What would change our mind: sustained convergence in the numbers over successive months, which would suggest the engines are settling on a shared authoritative core. A well-run study on consumer app surfaces rather than APIs showing materially higher overlap. Or the engines consolidating onto shared retrieval infrastructure, which would collapse the difference at a stroke. The harness runs daily, so we will see convergence if it starts, and we will publish it if it contradicts this page.
Published 29 July 2026. This page is re-checked as the measurement window extends.
Usually not. In Mode’s July 2026 study, three AI engines were asked an identical set of 53 questions every day for 17 days. Of the 1,378 domains they cited between them, only 26 (1.9%) were cited by all three, and 84.6% were cited by exactly one engine. On roughly a third of the questions where both OpenAI and Perplexity cited sources, they shared no domain at all. AI visibility is not one result. It is separate results per engine, and they disagree with each other far more than they agree.
In Mode’s measurement, Perplexity cited by a wide margin: an average of 15.0 sources on every single answer, 100% of the time. Anthropic’s Claude also averaged 15.0 sources but only carried citations on 57% of answers. OpenAI was the most selective, averaging 5.1 sources and citing on 34% of answers. More citations means more slots available, so an engine that cites 15 sources is mathematically easier to appear in than one that cites 5.
Sharply, on our sample. Perplexity leaned on community and social platforms: YouTube was 7.1% of its citations, LinkedIn 6.9% and Reddit 4.0%. Claude leaned on professional and reference sources: LinkedIn 9.9% and Wikipedia 5.9%, while citing Reddit and YouTube zero times. OpenAI leaned on official and industry sources such as government data, trade bodies and company registers (6.6% government or official), and cited Wikipedia and Reddit zero times. Three different diets, from the same questions.
The direction is well supported, the precise number needs care. It traces to Profound, which ran 100,000 identical prompts through ChatGPT and Perplexity and reported 11.0% shared domains, 37.4% ChatGPT-exclusive and 51.6% Perplexity-exclusive. A separate Whitehat SEO analysis of 118,000 answers across four platforms also reported 11%, but for “cited domains appearing across multiple platforms”, which is a different denominator to a two-engine pairwise overlap. Mode’s own replication on a smaller, UK-specific question set found 4.4% between OpenAI and Perplexity. Same direction, different magnitude. Treat 11% as an order of magnitude, not a constant.
Stop treating AI visibility as a single scoreboard. Check each engine separately, because being named by one tells you almost nothing about the others. Then work on the signals that travel across all of them rather than engine-specific tricks: a site machines can actually read, clear structured data about who you are, and genuine third-party mentions on the kinds of sources each engine favours. Mode runs this as a fixed-price AI Visibility Check, measured per engine and benchmarked against your three closest competitors.
We built this harness to measure ourselves, and it is the same measurement we run for clients. The AI Visibility Check tells you, per engine, whether AI mentions you, whether it can even read your site, and the top three fixes · benchmarked against your three closest competitors.
AI VISIBILITY CHECK · £95Honest by design: we measure signals, not guarantees. How GEO works.