How ChatGPT, Gemini, and Claude Read and Cite Your Content (2026)

We pointed our own measurement system at six AI engines and asked them the questions our buyers ask. The engines did not agree with each other. ChatGPT’s top source was Reddit, Google’s AI Overviews leaned on YouTube, and Perplexity cited Semrush first — same questions, three different “front pages.” Optimizing for “AI search” as one target means optimizing for an average that none of the six engines actually uses.

Our first-party study, not theory, is the foundation here. We ran roughly 27 real category prompts — “best AI SEO tools,” “how do I get cited in ChatGPT,” “what is generative engine optimization” — through ChatGPT, Microsoft Copilot, Google AI Overviews, Google AI Mode, Perplexity, and Gemini, and recorded which domains each engine cited. Below is what we found, why the engines diverge, and the per-engine playbook we use at Indexable to win the citation rather than just the crawl. Where an independent 2026 study by researcher Dan Petrovic (DEJAN) measured the same behaviors, we cite it as corroboration.

What did our cross-engine citation study actually find?

---

Our study found three distinct citation regimes, not six. The six AI surfaces we tested cluster into three families that source answers in fundamentally different ways:

The single most useful takeaway: the two Google surfaces behave almost identically to each other, ChatGPT and Copilot form a second pair, and Perplexity and Gemini form a third. A brand that wins one regime can be invisible in another. In our data, YouTube was the #1 cited source in Google AI Overviews (16 of our category responses cited it) yet did not crack ChatGPT’s top tier at all. arXiv showed up only in ChatGPT. Video is a Google-AI lever; research weighting is a ChatGPT lever. An engine-agnostic GEO strategy is a myth, and our own measurement is the proof.

Why does every AI engine cite a different #1 source?

---

Each engine runs a different retrieval-and-citation pipeline, so the same question lands on a different source set. Our study captured which sources each engine prefers; the DEJAN study measured how aggressively each engine converts what it retrieves into what it cites. Both line up.

The architecture differences are the durable insight: ChatGPT casts a wide net and cites a sliver, Gemini reads almost nothing yet cites nearly all of it, and Claude reads deeply and re-verifies its citations on a second pass (DEJAN, 2026). Treat the exact percentages as directional — they come from a single external study on a small query set — and treat the wide-versus-light-versus-deep behavior as stable enough to engineer for.

What did we find when we measured our own brand?

---

Before telling anyone how to get cited, we ran the measurement on ourselves — and the result was split. Our domain is cited in all six engines; our brand name barely registers. Those are two different outcomes, and most dashboards blur them into one.

The encouraging half: indexableai.com was cited in all six engines and ranked top-5 in four of them — #5 in ChatGPT, #2 in Perplexity, #3 in Gemini, #7 in Google AI Overviews. Two specific pages already earn ChatGPT citations: our AI-visibility page and our best-AI-SEO-agents guide. Our content is genuinely in the answer.

The uncomfortable half: when we measured brand-name share of voice against our tracked competitors, Indexable barely scored — near zero in ChatGPT and Google AI Overviews, 0.09 in Perplexity, 0.31 in Gemini — while Nightwatch and SearchAtlas were named again and again. Our pages are strong enough to be a source. Our brand is not yet strong enough to be a recommendation.

Cited as a source vs. named as a brand — what is the difference?

---

Being a source and being a recommendation are two different achievements that require two different kinds of work:

• Your page gets cited. The engine pulled a fact from your URL and credited it. Your content did its job.
• Your brand gets named. The engine writes “tools like Nightwatch and SearchAtlas…”. Your brand did its job.

Indexable is winning the first and losing the second, and we only know that because we measured both. The gap is widest exactly where the engines lean on third-party listicles and forums rather than first-party pages — which is most of our category. The question to ask is not “am I in the answer?” but “am I the source or am I the recommendation?” Our own data says most brands, ours included, are further along on one than the other.

Do AI engines dominate the search results in your category?

---

In our category, the AI Overview fires on every commercial query we tested. We ran twelve head keywords — “ai seo agents,” “geo tools,” “how to rank in chatgpt,” “ai visibility tracking,” “best ai seo software,” and more — through a SERP overview, and the Google AI Overview was present on 5 of 5 of the head money-keywords we checked in detail. The category is effectively 100% AI-Overview-saturated.

The composition of those AI Overviews is the actionable part. They are assembled from “we tested N tools” listicles, YouTube videos, and Reddit threads — and indexableai.com is not yet in the listicles that feed them. Reddit ranks organic #1–#4 on nearly every keyword we checked. This is direct evidence for why listicle-citation outreach is our top GEO priority: the money-keyword AI Overview is literally built from roundups we are not part of yet.

How does ChatGPT choose what to cite?

---

ChatGPT retrieves broadly, then filters hard through a narrow reading window. The DEJAN study measured it sourcing 39 pages for a single query and citing only 2 — a 5% citation rate — and evaluating each candidate through a 100–200 word sliding window rather than the full page. Our study shows what survives that filter: Reddit threads, research (arXiv), tech-press reviews, and definitional explainers.

The window is the whole game, and the rule we draw from it is our own: a citable claim has to be a self-contained, quotable unit inside about 200 words. A statistic set up in one paragraph and paid off three scrolls later never appears whole inside the window, so the claim loses the cut. Front-loading is not enough — the claim must be atomic: subject, claim, and number in one sentence. “Indexable cut customer time-to-first-byte 40%” survives the window; “It improved by 40%” does not, because the window may not contain what “it” refers to. On ChatGPT, structure beats volume because the reader is a narrow window, not a full crawl.

How does Gemini choose what to cite?

---

Gemini cites nearly everything it retrieves — close to a 1:1 ratio (DEJAN, 2026) — and it reads the least of any engine, leaning on snippets instead of full pages. Its retrieval set is small and binary: a page is either in it or invisible. In our study Gemini’s top sources were Reddit and Frase, with indexableai.com cited 3rd.

A snippet reader makes Gemini a first-chunk-or-nothing engine. The direct answer and the structured data in your opening section carry the entire decision, because Gemini does not give a down-page second chance. Lead with the answer, mark it up with JSON-LD schema, and assume nothing below the fold is read. One measurement trap: Gemini’s source URLs are redirect-wrapped (DEJAN, 2026), so its referrals are masked in GA4. A brand can be cited heavily by Gemini and see almost none of it in standard analytics, which is exactly why Gemini visibility has to be measured by probing the engine directly.

How does Claude choose what to cite?

---

Claude reads the most and commits the hardest. The DEJAN study found Claude pulling full pages, firing a single query, then re-analyzing its citations on a second pass — ending up citing about 64% of what it sourced, far more selective than Gemini and far more generous than ChatGPT.

Because Claude verifies twice, it rewards depth and internal consistency: comprehensive coverage, claims that are independently supported, and evidence that holds up on a re-read. Thin or self-contradictory pages that slip through a single-pass engine get caught on Claude’s second look. Claude is the hardest engine to fool and the most worth writing thoroughly for. If ChatGPT rewards atomic structure and Gemini rewards a strong first chunk, Claude rewards substance.

What is the metric that actually matters?

---

The metric is source→cite conversion: of the pages an engine retrieved for a query, how many did it cite — and was yours one of them? “We are getting retrieved” is not a result. ChatGPT retrieves about 20 pages to cite 1 (DEJAN, 2026), so retrieval is table stakes and citation is the product.

Source→cite conversion reframes the diagnosis. A brand that is retrieved but not cited has an extractability problem — its content is findable but not quotable — and the fix is structural. A brand that is never retrieved has an authority or relevance problem, and the fix is different. Most AI-visibility tools only report whether you were mentioned, so they cannot tell you which of those two problems you have, which means they cannot tell you what to do next. Separating “not retrieved” from “retrieved but not cited” is the single most useful thing a measurement system can do, and it is what Indexable was built to measure and then fix.

Does our own content pass the bar we recommend?

---

We held our own library to this standard, and most of it failed. We ran all 29 of our published articles through the same atomic-claim checker our Content Engineer uses as a pre-publish gate. Only 24% passed the 60% atomic-claim bar — 7 articles passed and 22 failed. The mean atomic-claim ratio across the library was 0.38, meaning fewer than four in ten of our citable sentences stand alone well enough to survive ChatGPT’s window. Eight articles scored 0.00, with their best claims fully buried — and they were mostly our narrative thought-leadership pieces, the ones written to be read by humans rather than extracted by machines.

Our 24% pass rate is the whole argument in miniature. Pages written to read well and pages written to get cited are not the same artifact, and even a team that knows the rule drifts from it the moment a piece turns essayistic. The fix is not better writing; it is a gate that checks every claim before publish.

What does this article look like through that gate?

---

We ran this article through the gate too, and it is the case study. On the structural checks it passes: an atomic-claim ratio of 0.78, a front-load ratio above target, fifteen headings (most in question form matching real queries), an answer-first opener, and tables and lists throughout. On two checks the first draft failed — and those failures are why this is the upgraded version:

• Zero structured data. The original draft shipped with no JSON-LD at all — no FAQPage schema, no BreadcrumbList — despite structured data being one of the largest citation levers in the research (FAQPage correlates with +45.6% and BreadcrumbList with +46.2% in AirOps’s 2026 Fan-Out analysis). This version ships both.
• Three pronoun-led claim openers. Three sentences opened with “That makes…”, “This reframes…”, and “It depends…”, each leaning on the sentence before it — exactly the dependency that fails ChatGPT’s window. All three are rewritten subject-first in this version.

Our own gate caught real defects in our own flagship — and you are reading the fixed page. The before-and-after is the dogfood: the same gate we sell, run on the work we publish.

How do you write content that gets cited across all three engines?

---

Write to the strictest reader for each job, and the others come for free. You can implement all four moves on a single page:

• For ChatGPT — atomic claims. Make every citable claim a standalone sentence (subject + claim + number) that survives a 100–200 word window. No claim that depends on a paragraph above it.
• For Gemini — win the first chunk. Put the direct answer and JSON-LD schema in the opening section; treat everything below as unread by the lightest reader.
• For Claude — depth that survives a second read. Cover the topic completely, attribute every data point to a source, and keep claims internally consistent.
• For all three — structure for extraction. Use question-format headings that match real queries, answer-first sections, tables, lists, and schema (FAQPage and BreadcrumbList each correlate with markedly higher citation rates in AirOps’s 2026 Fan-Out analysis).

Indexable’s Content Engineer enforces these as an automated pre-publish gate, including the atomic-claim check that flagged the three buried openers in this very article. Use this playbook on your next page, then re-check it against the gate before you publish.

How should you measure whether AI engines are citing you?

---

Measure per engine, measure both source and brand, and do not trust analytics to tell the whole story. ChatGPT passes trackable UTM-tagged links (DEJAN, 2026), so its referrals appear in GA4; Gemini’s links are redirect-wrapped, so its citations do not, and Gemini visibility has to come from direct probing. Our own study had to be run by probing the engines directly for exactly this reason.

Probe each engine the way a buyer would. Ask the questions your customers ask, repeat them several times, and record three things: whether you were retrieved, whether you were cited, and who beat you. Then track two numbers over time, per engine — your source→cite conversion and your brand-name share of voice — because, as our own data shows, you can be winning one and losing the other. You can do this manually in a spreadsheet or with a tool that probes the engines for you, but it has to be continuous: the engines change versions often, and a one-time audit goes stale the moment a model updates. For a hands-on method, see our ChatGPT visibility tracker guide, and for the broader playbook, how to rank in ChatGPT.

The honest caveat

---

Two kinds of data appear above, and they carry different confidence. Our cross-engine citation study and our own brand measurement are first-party, run on current models in June 2026 — directly observed, not modeled. The instrumented citation-rate figures (the 5% / ~100% / 64%) come from a single external study by Dan Petrovic (DEJAN) on a small query set, and model versions change quickly. Treat the architecture differences — tight versus wide versus deep, and the three citation regimes — as the durable insight, and the exact ratios as directional. Several things move the numbers: a model version update, a different query category (a developer question retrieves a different source set than a shopping question), and a brand’s own authority, which changes whether it gets retrieved at all. That is exactly why measurement has to be continuous rather than one-and-done — the only way to know your real source→cite conversion on today’s models is to measure it on today’s models, then keep measuring as they change.