AI Visibility · 8 August 2026

By Eddie Eastwood, Founder

ChatGPT vs Claude: How Often Each One Cites Your Business

A business that gets named by ChatGPT and never comes up in Claude has not failed at AI visibility. It has cleared one bar and not the other, and the two bars are not the same height or shape.

OtterlyAI's June 2026 citation study, tracking a continuously monitored set of SaaS and technology prompts, found that only 13% of the domains ChatGPT and Claude cited overlapped at all, and about 4.2% overlapped down at the individual URL level.^[1] Those are two different measurements on the same comparison: shared domains, and shared individual URLs. A business can clear one and miss the other with nothing about its content changing in between.

What matters is what each engine actually requires before it will put a business's name in an answer, and whether doing the work for one gets you most of the way to the other for free. It mostly does, but only from the Claude side. Building only for ChatGPT does not reliably satisfy Claude. This is the asymmetry the rest of the article works from, and it's also why it's worth showing you how to check where your own business stands, on both engines, this afternoon.

If you are weighing ChatGPT and Claude as tools for producing your marketing copy, that's a different question with a different answer, covered separately in Claude vs ChatGPT for SEO content. This article is about which engine cites your business once the content already exists, not which one should draft it.

One word, three decisions

"Citation" sounds like a single event: the engine either names your business or it doesn't. In practice an engine makes three separate decisions before a business's name appears in an answer, and most comparisons collapse all three into one.

First, retrieval: did the engine's index or live search even surface your pages for this query. Second, trust: given what it found, does the engine believe the page is accurate and specific enough to stand behind. Third, expression: does the engine state your business plainly, or does it hedge, naming a possibility while qualifying it, rather than recommending it outright.

Most comparisons of AI engines stop at retrieval, because it's the easiest thing to measure: did the name show up, yes or no. That misses the middle step entirely. A business can be retrieved by both engines and still get treated completely differently once each engine decides whether it trusts what it found.

Claude in particular will surface a business in its search results and then decline to recommend it plainly, offering a qualified mention instead: named, but hedged. That third state, named-but-hedged, does not exist in a binary named-or-not framework, and it is where a business that "sort of" shows up in AI answers has landed.

Where each engine actually looks

Once a business clears retrieval, what it gets cited for depends on where each engine is looking in the first place, and the two engines are not looking in the same places.

Claude's citation pattern, per OtterlyAI's June 2026 dataset, skews heavily toward brand and product domains: company, product and service websites accounted for 64.0% of all citations in the dataset, more than four times the next largest category, news and media, at 14.9%.^[2]

That doesn't mean Claude cites one dominant brand over and over. The publisher's own data guards against that read directly: the single most-cited brand domain accounted for under 3% of brand citations, and the top ten brand domains together made up only about 13% of that whole category. The 64% describes a preference for a type of source (a business's own site, its own documentation, its own product pages), not a concentration in a handful of winners.

That preference lines up with a claim ChatGPT-focused research has made about a different platform: Reddit is one of the most-cited social sources on engines such as Perplexity and ChatGPT, according to prior published research (see how often community sources get cited). Claude's numbers complicate that picture rather than confirm it. In the same June 2026 dataset, Claude cited Reddit zero times.

Social platforms as a category made up just 0.9% of all its citations, and within that thin slice, LinkedIn alone accounted for 43.1% of the social share.^[3]

That "zero" is scoped tightly: one month, one query domain, one dataset, and it is not a claim that Claude never cites Reddit under any circumstances. It is a measurement of what happened when a large, continuously tracked set of SaaS and technology prompts ran through Claude for four weeks. Within that scope, community-forum content essentially did not register, and a business's own site did the heavy lifting instead.

Both figures, the 64% brand share and the domain and URL overlap numbers above, come from the same publisher, the same one-month dataset, and the same SaaS-and-technology query set. They should be read as one coherent snapshot of that slice of Claude's behavior.

Why one states plainly and the other qualifies

The difference in what gets cited traces back to a difference in how each engine is built to retrieve and use what it finds, and both companies document the mechanism directly.

Anthropic's own documentation for Claude's web search tool describes a filtering step most comparisons never mention: "With basic web search, every search result is loaded into Claude's context window, and much of that content can be irrelevant to the request."

In its current implementation, Anthropic writes, "Claude instead writes and runs code that filters the results first, so only relevant content reaches the context window."^[4]

That's a documented design decision, not an inferred personality trait: a page has to survive an automated relevance check before Claude will even consider using it, on top of whatever judgment the model applies once it does.

Anthropic also documents when Claude reaches for the web at all: for "recent events, news, or announcements," "current prices, rates, scores, or statistics," and "information about specific organizations, people, or products that might have changed," and when it answers from what it already knows instead, for "established facts, math, science fundamentals" and similar stable knowledge.

OpenAI's crawler documentation describes a different kind of control, at the access layer rather than the relevance layer.

ChatGPT search runs on OAI-SearchBot, which is governed independently from GPTBot, the crawler used for model training: "a webmaster can allow OAI-SearchBot in order to appear in search results while disallowing GPTBot to indicate that crawled content should not be used for training OpenAI's generative AI foundation models."^[5]

A site that blocks OAI-SearchBot is explicit about the consequence: it "will not be shown in ChatGPT search answers, though can still appear as navigational links." OpenAI documents a separate agent, ChatGPT-User, for pages a user's own query causes ChatGPT to visit in the moment, and is equally direct that this agent "is not used to determine whether content may appear in Search."

Put the two together and the difference in behavior stops looking like temperament and starts looking like architecture. ChatGPT's search access is a permissions question decided mostly at the crawler level: allow OAI-SearchBot and your content is eligible to surface. Claude adds a content-quality gate after retrieval, filtering out what its process judges irrelevant before the model ever reasons over it. One system asks whether it's allowed to look. The other asks whether what it found is good enough to use, every time.

Build to the higher bar

That difference is why the asymmetry holds in the direction it does. If a page is specific enough, sourced enough, and checkable enough to survive Claude's relevance filtering and get cited plainly rather than hedged, it has already met a standard well above what OAI-SearchBot's crawl-and-index approach requires. A page that clears the higher bar clears the lower one as a byproduct.

The reverse doesn't hold. A page can be well-optimized for ChatGPT search (allowed in robots.txt, indexed, competitive for the query) and still read to Claude as thin: generic claims, no specifics an automated filter can verify, nothing that distinguishes it from a hundred similar pages. That page gets retrieved by both engines and cited plainly by neither, or cited by ChatGPT and hedged by Claude.

Ranking well on a traditional search surface does not transfer automatically to an AI answer engine either, a gap covered separately in Google SGE vs. ChatGPT Search. The same holds between AI engines: strength in one has to be earned again in the other.

That is why building to the strictest evidence bar first is the more durable rule. Content built only for the more permissive engine has no obvious next step when it stalls, because it was never built to survive the stricter filter.

Why a blended score hides the problem

It is tempting to average the two, call it "our AI visibility score" and track one number. A blended figure describes neither engine, for the same reason averaging a student's scores in two different subjects tells you nothing about either one.

The two engines draw from source pools that barely overlap, weight what they find differently, and apply a different bar before naming a business plainly. A single number that combines "named plainly by ChatGPT, hedged by Claude" and "cited by neither, but close on both" would report the same score for two situations that need completely different fixes.

This matters more than it sounds, because citation, click and revenue are three separate events even within a single engine, a distinction covered directly in Perplexity vs. Google for brand citation and referral traffic. Layering a second engine into a blended score compounds the same problem: it hides which specific gap is costing a business visibility, and a business cannot fix a gap it cannot see.

How to find your own citation rate, this afternoon

No published study measures how often any single business gets cited by ChatGPT versus Claude. The OtterlyAI dataset above measures where each engine's citations come from in aggregate, across thousands of domains, not how often it names one specific business. There is no shortcut around that: the only reliable way to know your own frequency is to run the check yourself, on your own real buyer questions, and count.

This isn't a comparison of how each engine finds your content in the first place. That mechanism is covered in ChatGPT vs. Perplexity: how each engine finds your content. It's a direct read on what each engine does once it has found you, and it takes about twenty minutes.

Write down 8 to 10 real questions a buyer asks before choosing a business like yours. Not your keywords: their words. Mix three types: a direct request ("who does [service] in [your city/niche]"), a comparison ("[you] vs [a named competitor]"), and a qualifying question a buyer asks before committing ("is [category] worth it," "what does [service] actually cost," "how do I know a [category] business is any good"). The comparison and qualifying questions matter more than the direct one; that's where a hedge shows up.

Ask every question in both ChatGPT and Claude. Open a new chat for each question, the "new chat" or "+" button in each tool, instead of asking all your questions in one running conversation. If the same conversation already named you once, it tends to keep naming you out of politeness rather than by re-deciding, which is not a real reading.

For each answer, record one of three outcomes, using these rules so the count stays consistent question to question:

• Named plainly. Your business is stated by name, as a recommendation, with a specific reason attached (a service, a location, a specialty). This still counts as plain if you are one of two or three named businesses, as long as the engine states you outright rather than qualifying the mention. If your name appears but the service area or specialty given is wrong, count it as plain but make a note: that's a content-accuracy problem to fix separately, not a citation problem.
• Named with a hedge. Your business appears, but qualified: "one option is," "you might also consider," "there are firms like," no clear recommendation attached, or you are one name buried in a longer list of many with no distinction given. If the answer names you but adds a caveat about whether it's confident, that is a hedge.
• Not named. A competitor is named and you are not, or the answer stays generic with no business named at all, on either side.

Some runs fit none of the three, for instance when the engine refuses to name any business at all. Leave those out of the count and note them separately rather than forcing them into a bucket. Keep the raw count per engine: out of 10 questions, how many landed in each of the three buckets, for ChatGPT and separately for Claude. That per-engine tally is your actual citation frequency, for your business, on the questions that matter to your buyers, which is the only frequency figure worth having.

Repeat the same 8 to 10 questions three or four times, spaced a few days apart over a couple of weeks. Both engines vary their answers between sessions and over time as retrieval and indexes update, so one pass tells you what happened once, not what happens generally. A pattern held across three or four runs is the number to trust; a single run is a snapshot, not a rate.

Read the pattern, not just the count. If a business lands mostly in named-plainly on both engines, the content is working: keep monitoring, don't rebuild. If it's named-plainly on ChatGPT but hedged or absent on Claude across every run, that's the asymmetry above showing up in practice: the content clears the more permissive bar and not the stricter one, and the fix below is on-domain work.

If it's absent on both, the problem is more likely retrieval than trust: check crawler access before anything else. The named-but-hedged outcome deserves the closest attention, on either engine: found, considered, and not quite trusted enough to recommend outright.

What to fix, and where

Once your tally shows which engine is the weak one, the fix depends on which engine it is, because the two gaps live in different places.

A ChatGPT gap is usually off-domain work. If ChatGPT isn't naming a business that ranks reasonably well and has decent content, the missing piece is often third-party validation the engine can find independently of the business's own site: coverage on review platforms, mentions in comparison content written by someone else, earned coverage in outlets an AI search index already treats as credible.

This is work that happens away from a business's own pages, building the kind of earned mentions on other sites that an engine can cross-reference rather than take on faith from a single source.

A Claude gap is usually on-domain work. If Claude is retrieving a business's pages but hedging instead of naming them plainly, the fix is rarely about getting more third parties to talk about the business. It's about whether the business's own pages give Claude's relevance filter something specific enough to pass through.

That means trading superlatives for specifics: not "the best solution for growing teams" but the actual capability, the actual number, the actual comparison a reader, or a filtering process, can check. Reference-grade product pages, comparison pages and how-it-works pages state facts plainly, with enough detail to be checked rather than taken on trust. That is what survives a filter built to strip out vague claims.

A business with gaps in both engines should start with the Claude gap. Fixing the specificity problem tends to improve the ChatGPT read as well, for the same reason building to the higher bar clears the lower one. Fixing only the ChatGPT-side gap, more third-party mentions, more coverage, does nothing for a page that Claude's filter is still discarding as too generic to cite.

Frequently asked questions

Does ChatGPT or Claude cite businesses more often?

No published study measures how often either engine names one specific business, and a single rate would not travel between them anyway, because the two engines cite from largely separate source pools. Showing up in one engine's answers is not a reliable predictor of showing up in the other's. The two-engine audit in this article is how to find out where your own business actually stands, question by question.

Do ChatGPT and Claude pull from the same sources?

Mostly not. OtterlyAI's June 2026 study found only 13% of the domains cited by ChatGPT and Claude overlapped, and about 4.2% overlapped at the individual URL level, across a continuously tracked set of SaaS and technology prompts.

Why does Claude hedge instead of naming a business plainly?

Claude's documented process filters search results for relevance before the model reasons over them, and applies its own judgment about whether a page is specific and verifiable enough to state confidently. A page that clears retrieval but reads as generic tends to get a qualified mention rather than a plain recommendation.

If I am cited by ChatGPT, will Claude cite me too?

Not automatically. Being cited by the more permissive engine says little about whether the stricter one will do the same, because the evidence bar is higher. The reverse holds more often: content built to satisfy Claude's evidence bar tends to also satisfy ChatGPT's.

How does crawler access affect whether Claude can cite me?

Anthropic documents that Claude reaches live pages through a web search tool rather than from training data alone, so anything blocked at the crawler level is outside what Claude can retrieve in the moment. Check your own server logs for the agents actually requesting your pages before writing a robots.txt rule for either engine.

Can I pay to be cited by ChatGPT or Claude?

No. Neither company documents a paid placement mechanism for citations in answers. OpenAI's crawler and search documentation describes citation eligibility purely in terms of crawl access and indexing; Anthropic describes citation as a function of its retrieval and relevance process. Citation is earned through what an engine finds and trusts, not purchased.

Is it worth building for Claude if most of my audience uses ChatGPT?

Yes. The work is the same work on both: specific, checkable, well-sourced pages. A business that builds to the stricter bar gives up nothing on the more permissive one, so serving Claude costs little beyond what serving ChatGPT already demands.

How often should I re-run the two-engine check?

Often enough to see a pattern rather than a single snapshot. Both engines vary between runs, so three or four checks spread over a couple of weeks tell you far more than one pass, and the check itself costs only the time it takes to ask the questions.

What does it mean if Claude names my business but qualifies what it says?

It means Claude retrieved and trusted the page enough to reference it, but the content didn't give the model's filtering process enough specific, checkable detail to state a confident recommendation. That's a signal to add specifics, real numbers, real comparisons, real capability detail, rather than to rewrite the page's tone.

Does the fact that Claude rarely cited Reddit in this study mean it never will?

No. The zero-Reddit finding holds within a specific scope: one month, SaaS and technology prompts, one dataset. It shows that community-forum content essentially didn't register in that slice of Claude's citations, not that Claude is built to exclude a source category outright.

Where to go from here

A free QBiz Leads AI visibility check tells you which side of the two-engine gap your business is actually on. It scans your website in about thirty seconds and returns a clear pass or fail on the key signals that decide whether AI tools can find and recommend your business.

Get your AI Visibility audit →