AI Visibility Research
AI Search Visibility Report 2026
ChatGPT, Gemini, and Perplexity rarely look at the same sources when they answer the same question. We analyzed 12,678 AI responses across 36 brands to see exactly how their source choices, page preferences, and recommendation patterns diverge.
Most marketing teams treat AI visibility as a single yes-or-no question: is my brand visible in ChatGPT? Our data says that question does not really hold up. When we tracked the same prompts across ChatGPT, Gemini, and Perplexity, each model consistently pulled from a different set of sources, favored different page types, and rewarded different content strategies. Optimizing for one model rarely transfers to the other two, and sometimes it barely helps at all.
This report walks through what the data actually shows: how much source overlap exists between models, where citations really come from, and which levers move a brand from invisible to recommended.
12,678
AI responses analyzed
1,368
Prompts tested
36
Brands tracked
20,005
Unique domains cited
How we did this
Methodology
This report combines two waves of Clamint's ongoing GEO (Generative Engine Optimization) source-citation study. The headline figures above reflect the most recent measurement wave: 36 brands, 1,368 unique prompts, and 12,678 responses collected from ChatGPT, Gemini, and Perplexity, citing 20,005 unique domains.
The category, page-type, and intent breakdowns further down draw on the full underlying analysis of the first 27-brand cohort: 1,079 unique prompts, 8,919 responses after removing test and duplicate records, 78,964 individual source citations, and 14,746 unique domains, collected between May 17, 2026 and August 19, 2026 and processed programmatically with Python and pandas.
Domains were sorted into 13 categories: brand's own site (1st party), marketplace and e-commerce, forum and UGC, review and comparison, news, encyclopedia, blog, corporate and official, competitor brand, competitor SaaS, service provider, and a catch-all "blog and other small-scale content" bucket for the long tail. The 300 most frequent domains were categorized by hand, and the remaining long tail was classified with rule-based heuristics using top-level domain, keywords, and subdomain patterns.
Page types (homepage, blog and guide, product and service, category, FAQ, video, or deep content) were inferred from URL path patterns. Prompt intent was classified with keyword dictionaries into commercial ("best X", "which one", "price"), informational ("what is X", "how does X work"), and navigational (prompts that name the brand directly).
Finding 1
Source overlap between models is almost zero
56% of ChatGPT and Gemini answer pairs, generated for the exact same prompt at the exact same time, do not share a single common source. Across the full dataset, the average overlap between any two models' source lists for the same prompt sits between 4% and 12%.
Average source overlap between model pairs, same prompt
Overlap measured as Jaccard similarity: the size of the intersection divided by the size of the union of each model's cited domains.
Gemini and Perplexity overlap the most, at 11.9%, most likely because both models lean heavily on live web retrieval. ChatGPT sits apart from both, sharing only around 5.2% of its sources with either one. In practice, this means a single generic SEO or GEO checklist will not move all three models at once. Each one needs its own targeted source strategy.
Finding 2
ChatGPT reads narrow. Gemini and Perplexity read wide.
On average, ChatGPT cites 4.25 unique domains per response. Gemini cites 12.89, and Perplexity cites 14.09, roughly three times as many. This ranking held across every brand in the comparison.
Average unique domains cited per response, by model
ChatGPT appears to concentrate on a small number of sources it treats as authoritative, mostly official brand pages, product pages, and a handful of comparison sites. Gemini and Perplexity spread their attention across a much wider pool, regularly pulling in blogs, forums, and marketplace listings that ChatGPT rarely touches.
Finding 3
Community content is now a distribution channel
Only 2.5% of ChatGPT answers cite at least one Reddit, YouTube, or forum source. For Gemini that share is 44%, and for Perplexity it is 52%, climbing to 62% specifically for decision-stage, comparison-style prompts.
Share of responses citing at least one community source (Reddit, YouTube, forums)
This lines up with the underlying category data: forum and UGC plus video and social sources make up just 1.2% to 1.3% of ChatGPT's citations, versus three to five times that share for Gemini and Perplexity. If your category has active Reddit threads, YouTube reviews, or forum discussions, Gemini and Perplexity are very likely already reading them, whether or not your brand is part of that conversation.
Finding 4
Being visible is not the same as being recommended
Gemini shows the widest range of brands, with a visibility score of 33.54%. But Perplexity is the most likely to place a mentioned brand near the top of its answer, doing so 62.08% of the time, compared with 44.85% for Gemini and 52.29% for ChatGPT.
Visibility score vs. share of mentions in the first 25% of the answer
Visibility score reflects how often a brand appears anywhere in a response. Prominence reflects how often it appears early, in the first quarter of the answer.
A single visibility score cannot tell these two situations apart. A brand can appear often but stay buried at the bottom of the answer, or appear rarely and still land in the spotlight almost every time it does. Judging AI visibility on one number hides which of these is actually happening to your brand.
Finding 5
Visibility is not stable from one day to the next
13% to 15%
Across 52 days of repeated measurement, brand visibility scores shifted by this much between consecutive checks, even when a model queried the exact same pool of sources. Models do not reliably pick the same brand twice from the same source pool. A single-date snapshot cannot tell you your real position. Only repeated measurement over time can.
Finding 6
Getting into the source set pays off, a lot
80% to 95%
This is the single strongest relationship in the entire dataset. Once your site becomes one of the sources a model actually cites for a prompt, the odds that it also becomes the brand the model recommends jump into this range. Being crawlable, indexable, and readable by AI systems is not a nice-to-have. It is the precondition for everything else in this report.
Deep dive
Where the citations actually come from
Across 78,964 source citations, blogs and other small, long-tail content sites account for 51.3% of everything the models cite. Marketplaces and e-commerce sites come next at 18.2%, nearly seven times the 2.7% share earned by brands' own official websites.
Share of citations by source category
The top 20 individual domains combined account for only 15% of all citations, and just 2 of those 20 are brand-owned sites. The rest are marketplaces, forums, or other third-party publishers. GEO visibility is not built on one dominant platform. It comes from a wide, fragmented ecosystem of small sources, most of which a brand does not control. A brand's own site is cited in at least one source in only 22.3% of responses overall, which means 77.7% of answers never mention the brand's own website at all.
Deep dive
Deep, specific pages get cited. Homepages do not.
Homepages account for only 9.5% of cited URLs, and category pages for a mere 0.6%. Blog and guide content (16.5%) plus deep, question-answering content pages (61.6%) make up nearly 80% of every citation.
Share of cited URLs by page type
There is also a model-specific pattern. ChatGPT cites homepages (18.0%) and product pages (13.1%) about twice as often as Gemini and Perplexity, while Gemini (17.9%) and Perplexity (18.7%) cite blog and guide content roughly two and a half times more than ChatGPT (7.2%) does. This tracks with finding two: ChatGPT trusts a small number of official pages, while Gemini and Perplexity trust a much broader library of independent guides.
Page-type preference, by model
Deep dive
Buying intent changes the mix, a little
59% of the prompts in this study were commercial in intent, things like "best X", "which one", or "price". The remaining 41% were informational, things like "what is X" or "how does X work".
Marketplace and comparison sites earn a noticeably bigger share of citations on commercial prompts than on informational ones. Brand-owned sites also do slightly better on commercial prompts, appearing in 23.7% of those responses versus 20.4% for informational ones. It is a modest effect, brand websites are still cited in less than a quarter of commercial answers, but it confirms that product and service pages carry a little more weight once someone is closer to a buying decision.
Deep dive
First-party visibility depends on the brand, not the industry
Brand-level first-party attribution ranged from 0% to 65.8% in the underlying dataset, with an overall average of 22.3%. That is an enormous spread for brands operating in similar categories.
First-party (brand's own site) attribution rate, selected brands
Five of the 27 brands we tracked had no first-party domain the models would reliably associate with them at all, no matching content ever surfaced. On the other end, the top-performing brands in the study (65.8% and 49.8% attribution) show what strong performance looks like: deep libraries of guide and FAQ content, multiple localized domains, and content built to directly answer high-volume questions. The gap between the best and worst performers is explained far more by content depth and technical readiness than by industry.
What to do about it
Five ways to act on this data
Build a model-specific content strategy
Cross-model source overlap is only 5% to 12%, so a single GEO checklist will not work for all three models. For ChatGPT, invest in a small number of strong, authoritative brand and product pages, plus homepage optimization. For Gemini and Perplexity, prioritize a wide footprint of third-party guides, comparisons, and forum presence.
Publish deep, question-first content, not more homepage or category pages
Homepages and category pages combine for barely 10% of citations. Every high-volume "best X", "how to choose X", or "X vs Y" query deserves its own dedicated article. FAQ content is a cheap, underused opportunity at under 1% share, structured FAQ markup has real room to grow.
Actively manage marketplace and comparison-site listings
Marketplaces alone outweigh brand-owned sites by roughly seven to one. Keep product listings on relevant marketplaces complete, current, and detailed with specs, FAQs, and comparison tables. Regularly audit how your brand is represented on independent review and comparison sites.
Invest in community and forum reputation
Forum and UGC citations are three to four times more common in Gemini and Perplexity than in ChatGPT. Monitor brand mentions on Reddit and complaint or review platforms: unanswered negative threads surface directly in AI answers. Encourage genuine customer conversation rather than incentivized reviews, which models appear to treat as a lower-trust signal.
Strengthen first-party visibility deliberately, brand by brand
The gap between the best (65.8%) and worst (0%) performers in our data is not explained by industry, it is explained by content depth and technical readiness. Start with crawlability, structured data, and page speed. Confirm that AI crawlers such as GPTBot, Google-Extended, and PerplexityBot are not blocked in robots.txt, then build direct-answer content for your highest-value queries.
If your team is optimizing for "AI search" as one undifferentiated channel, this data suggests you are solving the wrong problem. ChatGPT, Gemini, and Perplexity are three different discovery surfaces, each with its own sources, its own preferred page types, and its own tolerance for third-party noise. The brands with the widest visibility in this study did not get there with a single tactic. They showed up, consistently, in the specific places each model was already reading.