Foglift puts cross-engine AI citation overlap at 0.094 Jaccard while brand agreement runs 85-98%. Here is what that split means for B2B AI visibility reporting.
Five AI engines answered the same buyer questions. They shared 9% of their sources and 90% of their brand picks. That gap is the most useful number in AI search right now, and most reporting hides it.
AI citation overlap measures how often two AI engines cite the same source for the same prompt. Foglift's Q3 2026 benchmark ran 75 buyer questions across five engines, logged 375 answers, and counted 1,510 distinct cited domains. Mean pairwise overlap came out at 0.094 on the Jaccard scale. A citation on Perplexity buys you close to nothing on ChatGPT, while the brand recommendation carries across engines at 85% to 98% agreement.
Here is what the newest data shows and what to change in your reporting this month.
AI citation overlap is the share of cited sources that two engines have in common on the same prompt. Researchers measure it with Jaccard similarity: domains cited by both engines divided by all unique domains cited by either. A score of 1.0 means identical source lists. A score of 0 means nothing shared.
Three 2026 datasets put the real number between 0.09 and 0.24. Foglift's Q3 2026 panel of 75 buyer questions across five engines reported a mean pairwise Jaccard of 0.094 across 375 answers and 1,510 domains. Writesonic ran 161,286 prompts across ChatGPT, Gemini, Perplexity, and Google AI Overviews in May and June 2026 and found every engine pair between 0.119 and 0.237. SurfacedBy logged 127,198 citations from about 16,400 answers between March 29 and June 27, 2026, and found 2.7% of cited domains appeared on all five engines.
No study puts overlap above a quarter. Track this metric for your brand → nobori.ai
Seven in ten cited domains appear on exactly one engine. SurfacedBy's 11,647-domain sample breaks down at 69.6% cited by one engine, 16.3% by two, 7.4% by three, 4.1% by four, and 2.7% by all five. That last group is 309 domains out of 11,647. Writesonic's four-engine cut lands in the same place: 72% to 73% of cited domains appear on one engine, and 3.8% appear on all four.
The pairwise numbers show where the fault lines run.
| Engine pair | Jaccard overlap | Prompts measured |
|---|---|---|
| Perplexity and AI Overviews | 0.237 | 126,299 |
| Gemini and AI Overviews | 0.216 | 87,558 |
| Gemini and Perplexity | 0.174 | 82,221 |
| ChatGPT and Perplexity | 0.130 | 119,513 |
| ChatGPT and AI Overviews | 0.126 | 128,623 |
| ChatGPT and Gemini | 0.119 | 88,969 |
Source: Writesonic, July 2026, domain-level matching. ChatGPT sits at the bottom of every pair it appears in, with an average of 0.125 against the other three.
Brand agreement runs far higher than source agreement. Foglift tested 62 queries where all five engines returned an answer and measured both layers. Brand-mention agreement between engine pairs ranged from 85.5% to 98.4%. Source overlap on the same pairs dropped as low as 0.027.
Gemini and Google AI Overviews agreed on brands 98.4% of the time at 0.643 source overlap. Gemini and Perplexity agreed 96.8% at 0.197. Google AI Overviews and Perplexity agreed 95.2%. ChatGPT and Claude agreed 85.5% at 0.027.
Source: Foglift, June 2026, 1,373 AI answers. ChatGPT and Claude share 2.7% of their cited domains and still name the same brands most of the time. The engines read different shelves and reach the same verdict about who to trust.
BrightEdge's cross-engine analysis shows the same spread between layers: citation source overlap across engine pairs ranged from 16% to 59%, a 43-point spread, while brand mention overlap ranged from 36% to 55%, a 19-point spread.
Citation volume differs by a factor of three, which changes your odds before content quality enters the picture. SurfacedBy's five-engine sample recorded Gemini at 11.0 sources per answer, Perplexity at 8.6, Google AI Mode at 7.8, Claude at 6.8, and ChatGPT at 3.7.
A page that ranks eighth-best for a prompt has a real shot on Gemini and almost none on ChatGPT, where the answer names three or four sources total. Teams that report a single "citation rate" across engines average a 3-slot competition against an 11-slot one.
Volume also moves inside a single engine within weeks. SurfacedBy's August benchmark found ChatGPT sources per answer rising from roughly 10 to 12 before August 8 to 12 to 15 after, measured on live-interface answers rather than the API. Same engine, same prompts, different arithmetic 11 days apart. Our breakdown of AI citation half-life covers how fast the source list itself rotates underneath those counts.
Each engine has a source personality, and video exposes it faster than any other format. SurfacedBy's five-engine sample measured the share of each engine's citations going to YouTube: Google AI Mode 11.2%, Perplexity 8.8%, Gemini 2.2%, ChatGPT 1.6%, and Claude 0.02%.
| Engine | Sources per answer | Citations to YouTube |
|---|---|---|
| Gemini | 11.0 | 2.2% |
| Perplexity | 8.6 | 8.8% |
| Google AI Mode | 7.8 | 11.2% |
| Claude | 6.8 | 0.02% |
| ChatGPT | 3.7 | 1.6% |
Source: SurfacedBy, 127,198 citations across five engines, March to June 2026.
BrightEdge frames the same split by authority preference, reporting Gemini favoring institutional sources over user-generated content at a 130:1 ratio while Google AI Overviews sends 17.5% of citations to UGC. BrightEdge also found Perplexity naming 86% of the brands in an answer by position five, which puts a premium on early placement in its source list.
The same domain carries wildly different weight per engine. Ahrefs Brand Radar's September 2, 2026 snapshots, compiled by Obsurfable, put YouTube at 22.9% mention share on Google AI Overviews and 2.5% on ChatGPT. That is a 9x gap on one platform.
| Domain | ChatGPT | AI Mode | AI Overviews | Perplexity |
|---|---|---|---|---|
| 16.8% | 17.9% | 18.5% | 21.6% | |
| YouTube | 2.5% | 17.8% | 22.9% | 20.8% |
| Wikipedia | 7.0% | 4.0% | 4.0% | 6.3% |
| google.com | not in top 50 | 12.5% | 8.8% | not in top 50 |
| Consumer Reports | 3.7% | 0.5% | 0.4% | 1.3% |
Source: Ahrefs Brand Radar, September 2, 2026 snapshots (US, all topics), via Obsurfable. Mention share counts each domain's citations as a percentage of the top 50 sources on that assistant.
Two structural notes for B2B teams. Consumer Reports and Forbes carry 4x the weight on ChatGPT that they carry on Google AI Overviews. And google.com at 12.5% on AI Mode means your Business Profile and product panels are citation infrastructure, not local SEO hygiene.
Reddit fell off ChatGPT Search citations in two steps in the second week of August 2026. Promptwatch, reported by Search Engine Land on August 19, measured Reddit at 3.83% of ChatGPT Search citations from July 18 through August 7, then below 1% on August 14, averaging 0.52% through August 17. That is an 86.4% relative drop. Qwairy's separate corpus tracked a steeper fall, from 2.05% on August 1 to 0.07% on August 14.
The mechanism looks like retrieval policy, not reputation. Promptwatch's fan-out logs show background queries using the site: operator jumping from about 0.37% to 16.8% in a single day on August 8, then holding. Qwairy tracked the same operator at 9.7% on August 8, 20.4% on August 13, and 23% to 24% from August 14 to 16. Those scoped queries target brand domains, regulators, and institutional sites. Reddit cannot win a query pinned to a named site.
Government and regulatory pages absorbed most of it. SurfacedBy logged 2,044 ChatGPT answers in two windows, 1,434 between July 21 and August 7 and 610 between August 12 and 19, using the same prompts and brands in both runs.
| Source type | Before (1,434 answers) | After (610 answers) |
|---|---|---|
| 22.9% | 3.3% | |
| Government and regulatory | 11.9% | 23.9% |
| Review platforms | 3.1% | 4.1% |
| YouTube | 0.8% | 1.3% |
| Wikipedia | 1.5% | 1.5% |
Source: SurfacedBy, August 20, 2026. Percentages count answers containing at least one link of that type, not share of all links.
Review sites and video moved on numbers too small to trust. Wikipedia sat still. ChatGPT traded community opinion for official sources, and in regulated B2B categories those pages belong to agencies you cannot publish on.
No. The collapse was ChatGPT-specific, which is the whole argument for engine-separated reporting. SurfacedBy ran the same prompts and brands across five assistants in both August windows and found Google AI Mode easing from 47.2% to 34.2% of answers citing Reddit, Gemini from 42.1% to 27.9%, and Perplexity moving the other way, from 54.4% to 67.6%. Claude sat at 0% in both windows across 684 answers.
Promptwatch found no comparable single-day break on Google surfaces either, recording AI Overviews drifting from about 2.5% in early July to 2.1% in August, and AI Mode from 2.22% to 1.54%. Trackerly added a nuance worth logging: Reddit's share of pages ChatGPT consulted held between 25% and 34% from August 13 to 20 while citations collapsed. If that holds, the model still reads the thread and footnotes a .gov page instead. Reddit also concentrates at the decision stage on Google surfaces: EMGI ran 1,486 SaaS buying queries and found Google AI Overviews citing Reddit on 2.5% of top-of-funnel queries against 20.1% of bottom-of-funnel ones.
September 2026 marked the first Brand Radar month where Reddit outranked YouTube on Perplexity. Reddit rose to 21.6% mention share while YouTube fell from 31.2% in August to 20.8%. Perplexity stopped being a video-first engine inside four weeks.
The rest of Perplexity's table rotated toward health and editorial sources. Healthline climbed three places to 2.6%, the New York Times jumped 28 places into the top 10 at 1.6%, and Good Housekeeping rose 11 places. TikTok Shop fell 35 places. Automotive domains dropped: Edmunds down 14, Kelley Blue Book down 17, Cars.com down 23.
Anyone who built a Perplexity program on July advice to treat YouTube as the primary lever spent a quarter optimizing a source that lost a third of its weight. Monthly leaderboards lag mid-month policy changes.
Claude behaves like a separate channel. SurfacedBy's five-engine sample recorded Claude sending 0.02% of its citations to YouTube and 0.01% to Reddit while still listing 6.8 sources per answer. It pulls from documentation, vendor pages, and prestige editorial instead.
Part of the reason sits in infrastructure. ChatGPT retrieves through Bing and OAI-SearchBot, Claude through Brave Search, and Gemini through Google's index. Different indexes produce different candidate sets before any ranking logic runs.
Claude also names brands more often than the others. SurfacedBy logged about 26,400 answers across five assistants and found brands cited or named in 16% of them, with Claude highest at 20.3%, Perplexity at 17.5%, Gemini at 16.6%, and ChatGPT and AI Mode tied at 14.1%. For B2B teams whose buyers use Claude, a YouTube and Reddit program is the wrong primary bet.
A blended score averages source ecosystems that share less than a quarter of their inputs. The average describes no engine your buyer opens.
Run the math on a five-engine dashboard. If ChatGPT names 3.7 sources and Gemini names 11.0, a brand cited on Gemini and absent from ChatGPT can post a healthy blended number while losing every ChatGPT answer. The same flattening hid the August Reddit collapse: a program that looked stable on "Reddit share of AI citations" on August 7 was a ChatGPT-specific crash by August 14 while Google drifted at a tenth of the speed.
Volatility compounds the problem. SurfacedBy measured how often cited sources change between repeated checks on the same question and found Google AI Mode at 68%, Gemini 48%, Claude 45%, ChatGPT 35%, and Perplexity 25%. Ahrefs puts the odds of an AI Overview changing between observations near 70%. Writesonic's stability study across 631,999 prompts found 52% of ChatGPT's number-one brand positions rotating on the same prompt. Averaging five unstable series produces one number that moves for reasons you cannot name. See our data on the B2B AI visibility gap for what per-engine citation rates look like at benchmark scale.
Fragmentation cuts in your favor at the bottom of the distribution. SurfacedBy found 43% of cited domains earned exactly one citation, with the top 10 domains taking 20.6% of citations and the top 100 taking 42%. Geonimo's 2.1 million citation analysis put 73.5% of citations outside the top 100 domains. Foglift's Q3 panel found 75% of the combined top-25 domains exclusive to one engine.
SurfacedBy's commercial-intent sample also breaks the Reddit-and-Wikipedia consensus story. In that corpus Reddit took 1.8% of all citations and Wikipedia under 0.6%, while vendor pages, documentation, and long-tail category sites took 90.6%.
You do not need to outrank Reddit. You need the clearest page on the narrow question your buyer asks, because engines fan out past the mega-domains and grab whatever specific thing exists. Our analysis of product pages versus blog posts shows which page types win that fan-out.
Brand-level signals transfer across engines. Source-level wins do not. Hexagon analyzed 50,000 AI shopping citations and found 12% of brands captured more than 80% of all recommendations. The strongest predictor was high-authority editorial mentions at 0.74 correlation, followed by review volume, where brands with 500 or more reviews earned 4.2x the citations of brands under 50, and semantic consistency across surfaces at a 2.4x lift.
A University of Toronto study of 215 commercial prompts found ChatGPT and Claude disagreed on brand recommendations 67% of the time, then agreed on the reason for excluding a brand 95.1% of the time. The failure modes are shared: the brand never entered the model's view, or it entered and went unmentioned, or it got mentioned and not recommended. Fixing the failure mode moves both engines at once, which the researchers confirmed with their own interventions.
Editorial breadth beats editorial depth here, since each engine draws from a different publisher pool. Our guide to off-page AEO covers how to build that breadth.
Split every metric by engine, then add one cross-engine layer for brand presence. Five columns, not one score.
Report four things per engine. First, citation presence: did the engine cite your domain on this prompt. Second, brand mention: did it name you, cited or not. Third, source-type mix behind the answers that mention you, sorted into owned, review, editorial, community, video, and institutional. Fourth, share of weeks present on a fixed prompt set.
Then add the number that survives engine changes: brand agreement. Count the share of your priority prompts where you appear on at least four of five engines. Foglift's data says the ceiling on that number is high, between 85.5% and 98.4% for brands above the consensus floor. MaxAEO's 1,500-prompt study across 14 categories found engines share an average of 38% of their top five brand picks, so roughly two consensus slots exist per question. Those slots go to the same brands across queries in a category.
Start with a concentration audit, because concentration is the risk the August data exposed. Pull the sources behind every AI answer that mentions you, sort them by type, and find the single source type carrying more than 30% of your visibility. That is your exposure when the next retrieval change lands.
Then run these five steps.
Overlap has no reason to rise. Each engine now runs its own index, its own fan-out policy, and its own source preferences, and the September data shows those preferences moving on a monthly clock in different directions. Perplexity swapped its top source. ChatGPT rotated from dictionaries toward health and .gov. Google kept raising its own properties, with google.com reaching 12.5% on AI Mode.
The brand layer is the part that compounds. Engines already treat you as one entity, agreeing on brand recommendations 85% to 98% of the time while sharing under a quarter of their sources. Programs built on entity consistency, editorial breadth, and review depth clear the consensus floor and stay cleared through retrieval changes. Programs built on one source type get re-baselined every August. Our work on entity optimization for AI search covers the mechanics of that floor.
What is a normal AI citation overlap score? Published 2026 studies put pairwise Jaccard overlap between 0.09 and 0.24 at the domain level. Foglift's Q3 2026 mean was 0.094 across five engines; Writesonic's range across six engine pairs was 0.119 to 0.237. Exact-URL matching produces lower scores than domain matching.
Which two engines cite the most similar sources? Perplexity and Google AI Overviews at 0.237 in Writesonic's July 2026 study, both live-retrieval engines pulling from overlapping web indexes. Gemini and Google AI Overviews reached 0.643 in Foglift's smaller June panel.
Which engine is the biggest outlier? ChatGPT sits lowest in every pair it appears in, averaging 0.125 against Gemini, Perplexity, and AI Overviews. Claude runs a separate source personality, sending 0.02% of citations to YouTube and 0.01% to Reddit.
Ready to see where you stand in AI search?
Nobori tracks your brand's visibility across ChatGPT, Gemini, Perplexity, Google AI Overviews, and Claude, with each engine reported on its own. See who's getting cited instead of you, which source types carry your presence, and what to fix.
Get daily AI visibility alerts for your brand → nobori.ai
