Mack McConnell
A marketing team we worked with had a good quarter on AI search. Their ChatGPT mention rate climbed, Perplexity started citing their comparison page, and the monthly report finally had a line going up and to the right. Then someone on the executive team opened Claude, asked the same buying question the team had been tracking for months, and got an answer that named three competitors and not them. No citations, no links, nothing to appeal. The program was working everywhere except the one place the CEO happened to check.
That gap is common enough that we treat it as a separate workstream now. At Geostar we run visibility programs across all the major engines, and Claude is consistently the one that refuses to move on the same inputs that shift the others. Teams read that as a measurement problem or a content problem. It is usually neither. Claude is running on a different mechanism, and often on a different clock, than the engine the rest of the program was built around.
This piece covers what that mechanism actually is, what the available data says about who Claude cites, and where we would spend a constrained quarter if Claude specifically is the gap. Some of the numbers here come from vendor datasets rather than neutral research, so we have named the source of every figure and flagged what is not independently confirmed.
Why Claude Behaves Differently From ChatGPT and Perplexity
Claude cites brands less often than its peers, and by a wide margin. The GEO vendor Boring Marketing ran 14,298 buyer-question checks across four engines and reported a brand citation rate of 7.6% for Claude, against roughly 12.5% for ChatGPT, 17.6% for Perplexity, and 17.8% for Gemini. Those figures come from a tool vendor's own benchmark rather than an independent study, so treat the exact percentages as directional. The ordering, though, matches what we see in client tracking: the same prompt set that produces steady ChatGPT and Perplexity mentions produces a fraction of that volume in Claude.
The selectivity has a shape to it. A cross-sectional analysis of 10,038 Claude health citations, published as an arXiv preprint by researchers at York College CUNY and Columbia's Teachers College, found that established institutional sources accounted for 97.8% of citations. Within a deeper stratified sample of 542 of those sources, the median cited-domain Domain Authority was 92 [1]. That study covers consumer health queries only, which is a vertical where deference to institutional sources is both expected and appropriate, so it does not generalize cleanly to software or retail questions. It does tell you what Claude's retrieval prefers when it has a strong institutional option available.
Put the two findings side by side and the behavioral difference gets concrete:
| Engine | Brand citation rate, Boring Marketing benchmark | Practical implication for a visibility program |
|---|---|---|
| Claude | 7.6% | Fewest openings per prompt set, heaviest weighting toward established sources |
| ChatGPT | 12.5% | Middle ground, responsive to both owned and third-party work |
| Perplexity | 17.6% | Highest citation volume, most responsive to fresh retrievable content |
| Gemini | 17.8% | Highest citation volume, closely tied to conventional search signals |
Every figure in that table comes from a single vendor benchmark, so read the column as a ranking rather than a set of precise rates. The practical consequence is that tactics do not transfer cleanly between engines. Our ChatGPT guide and Perplexity guide both lean on fast-moving retrieval and fresh content, because both engines reward it. Claude rewards it less. A program tuned for citation velocity on one engine can run for two quarters against Claude and produce almost nothing, not because the work was bad but because it was aimed at a retrieval layer that is weighted differently.
The Two Distinct Ways Claude Can Recommend a Brand
Most confusion about Claude comes from treating it as one system. It is two, and they respond to completely different work.
Recall: what Claude already believes about you
In ordinary conversation, Claude answers many questions without searching anything. No search indicator appears, no citations attach, and your brand either surfaces or it does not based on how the open web described you when the training corpus was assembled.
This is the path that produces the scenario at the top of this article. An executive asks a plain buying question, gets a plain answer naming three vendors, and nothing about that exchange is traceable to a page anyone can go fix. Recall moves on Anthropic's model release cadence, which means work you do this month cannot surface until a future model ships. It is also the path most visibility tooling under-reports, because a response with no citations gives a tracker nothing to parse beyond the presence or absence of a brand name.
Retrieval: what Claude can find and quote right now
Anthropic's support documentation describes the second path directly. When a question benefits from current information, Claude invokes a search tool, processes multiple sources, and returns a conversational response with direct citations, source links, and relevant quotes [2]. Web search has been globally available on all Claude plans since the May 2025 update, and Anthropic states that citations appear whenever web information is incorporated into a response [3].
Retrieval moves in days rather than model generations. The catch is that it only opens on queries where Claude decides a search is warranted, and that decision is not something you control. Stable, evergreen, definitional questions are frequently answered from training data alone. The narrower and more current the question, the more likely a search runs and a citation slot opens.
The distinction between the two matters more than any individual tactic:
- Recall is unattributed, appears in default chat, and is influenced by the aggregate description of your brand across the web at corpus-build time.
- Retrieval is attributed, appears only on search-triggered queries, and is influenced by what is findable and quotable right now.
- A prompt that reliably triggers search for one phrasing may answer from training data on another, which is why two testers can get contradictory results from the same question.
We covered the retrieval side in more depth in our earlier piece on Claude's web search and what it means for brands. The point worth repeating here is that measuring the two together produces a number that means nothing. A brand can hold steady on retrieval while its recall position quietly erodes, and a single blended "Claude visibility" score will show neither.
What Makes a Page Quotable Enough to Survive Claude's Retrieval
Before the tactics, a caveat on the plumbing. Multiple GEO vendors, including BrightEdge, report that Claude's web search runs on Brave Search rather than Google or Bing. The claim traces back to Anthropic's own subprocessor list, which added Brave Search on March 19, 2025, and to independent write-ups that surfaced the change at the time. We have not found that confirmed on any Anthropic-owned page. The only backend Anthropic names explicitly in its own support documentation is Bing, and only for image search results, which is a narrower claim than the one circulating. Treat the Brave attribution as widely reported third-party reporting, not vendor-confirmed fact, and do not build a strategy that depends on it being true.
What is documented is how citations attach. Anthropic's web search tool documentation specifies that each citation carries a URL, a title, and a cited_text span capped at roughly 150 characters [4]. That constraint is the most useful structural fact available about Claude retrieval. The unit Claude extracts is not your page, and not your section. It is a short passage that has to stand on its own once it is lifted out of everything around it.
Writing for that constraint means changing what a paragraph is doing:
| Trait | What it looks like on the page | Why it survives extraction |
|---|---|---|
| Direct answer placement | The first sentence after a heading answers the heading's implied question outright | The extractable span sits where the retriever is most likely to look |
| Atomic claims | One fact per sentence, with the subject named rather than pronouned | A lifted sentence still identifies who and what it is about |
| Named sources | Figures attributed inline to the organization that produced them | The passage carries its own credibility without the surrounding page |
| Clean semantic markup | Real headings, lists, and tables rather than styled divs | Structure is parseable rather than inferred from visual formatting |
| Explicit comparison | Options and attributes laid out rather than described in prose | Comparison questions have something directly quotable to return |
None of that is exotic, and it is the same discipline that made a page useful to a human skimmer. The difference is the failure mode. A human skimmer who lands mid-paragraph can scroll up for context. A retriever working on a 150-character span cannot.
One clarification worth making, because almost nobody in this space makes it. Anthropic also publishes a Citations API for developers, which returns source references from documents an integrator has uploaded or otherwise provided directly into their own application [5]. It is a retrieval-augmented generation feature, not a consumer-facing one. It has nothing to do with whether Claude mentions your brand in a consumer conversation, and no amount of work on your website affects it. If you have seen the Citations API cited as a lever for organic visibility, that advice was mixing up two unrelated products.
Why the Larger Lever Sits Off Your Own Website
Here is the number that reorders most Claude roadmaps. The GEO vendor Omnia, working from its own citation database of nearly 500,000 Claude citations tracked across more than 37,000 answers during the first two weeks of its Claude tracking in 2026, reports that only 6.7% of Claude citations point to a brand's own domain. The remaining 93.3% go to third-party sources, and 90.1% of those are editorial and independent web content. This is Omnia's first-party dataset rather than independent research, so the precision is worth discounting. The magnitude is hard to explain away.
Read alongside the health-citation study's institutional skew, the picture is consistent. Claude is disproportionately quoting sources that are not you, about you. Which means the ceiling on any Claude program built entirely from owned-content work is low, no matter how well executed the content is.
The third-party work that compounds falls into three categories:
- Cross-source consistency. Every place that describes your product should describe it the same way: same category language, same core capability, same positioning. Directory listings, review platforms, partner pages, and old press coverage that contradict your current positioning actively hurt, because they hand the model competing descriptions to reconcile.
- Earned editorial coverage. Independent writers publishing on established domains produce exactly the source profile the citation data favors. This is slow, it does not scale on a content calendar, and it is the single highest-leverage line item for Claude specifically.
- Presence on already-cited pages. Roundups and comparison articles that Claude already pulls from are the fastest available route in. Finding which pages those are is a tracking exercise, and being absent from a page a model already trusts is a more addressable problem than trying to make your own domain trustworthy enough to displace it.
That third workstream is where we spend a meaningful share of client hours, because it is the one that converts research directly into placement. Teams that treat earned coverage as a PR line item disconnected from search tend to under-fund it, and Claude is the engine where that under-funding shows up first.
The uncomfortable part of this finding is that it inverts how most content budgets are built. A typical AI search program allocates the majority of its spend to publishing on the owned domain, which is the destination for well under a tenth of the citations in Omnia's Claude dataset. The owned-content work still matters, because a page that cannot be quoted cleanly will not be quoted when a retriever does land on it. But it is the smaller of the two levers, and treating it as the whole program is the most common reason a Claude number sits flat through a quarter of genuine effort.
Sequencing the Work: What to Do First This Quarter
Most teams asking about Claude are not starting a Claude program from zero. They are deciding how much of an existing budget to move, and how much to leave with the engines already producing. Here is the order we use.
- Audit how the web currently describes you, before touching your own site. Pull the top sources for your category questions and check what each says about your product. Contradictions between them are the cheapest fix available and the one with the most direct effect on the recall path.
- Fix quotability on the pages that already earn high-intent traffic. Not the whole site. The handful of pages that answer the questions you want to be the answer to. Direct answers after headings, atomic claims, named figures, real markup.
- Fund earned coverage as an ongoing line, not a campaign. Given the third-party skew in the citation data, this is where the ceiling lifts. It needs to run continuously, because a single burst of coverage ages out of relevance while a steady stream compounds.
- Test recall and retrieval separately, and report them separately. Run your prompt set once in default chat and once with search explicitly requested. Anthropic's own documentation notes that adding "search the web" to a prompt is how a user forces the search path [2], which makes the paired run straightforward to standardize. Log which mode answered, whether your brand appeared, and which domains were cited when sources were shown. Blending the two produces a number that cannot be acted on.
- Leave the working engines alone. If ChatGPT and Perplexity programs are producing, do not cannibalize them to chase Claude. The third-party work in steps one and three lifts all three engines. The Claude-specific spend is the earned-coverage acceleration, not a full program rebuild.
For teams that want the broader framework these steps sit inside, our complete guide to generative engine optimization covers the measurement and prioritization layer in full.
Frequently Asked Questions
Does ranking well in Google help with Claude?
Indirectly, and less than you would expect. Anthropic has not confirmed a search backend for general web queries, and the vendor reporting that circulates points to Brave rather than Google. Strong Google performance usually correlates with the things Claude does favor, like established domains and consistent third-party coverage, so it is a decent proxy but not a mechanism.
Does Claude always cite its sources?
Only when web search runs. Anthropic states that citations appear whenever web information is incorporated into a response [3], which is the same as saying that when no search runs, no citations appear. Stable, evergreen questions are frequently answered from training data alone, with no sources shown.
How do I test whether Claude knows my brand?
Run each prompt twice. Once normally, and once with an explicit instruction to search the web. If your brand appears in the second run but not the first, you have a retrieval win and a recall gap. If it appears in neither, the third-party work in the sequencing section is your starting point.
Does any of this change when a new Claude model ships?
The recall path does. That is the only time the training-corpus picture updates, which is why work aimed at recall has a lag measured in model releases rather than weeks. The retrieval path is unaffected by model releases and responds to changes in what is findable and quotable now.
Is the developer Citations API relevant to my visibility?
No. It returns references from documents an integrator uploads into their own application [5]. It is a separate product from Claude's organic web citation behavior, and nothing you publish affects it.
Where This Leaves You
The two clocks are the thing to hold onto. Retrieval responds within weeks to work on quotability and to earned placements on pages Claude already reads. Recall responds on Anthropic's release schedule, to the aggregate story the open web tells about you, and there is no shortcut through it. A quarter spent making your own pages more extractable will move the first and barely touch the second, which is the specific reason so many well-run programs stall on this engine.
If Claude is the gap in an otherwise healthy AI search program, the diagnosis is usually available in an afternoon of prompt testing plus an audit of how third-party sources currently describe you. Book a free audit and we will run both, and tell you which of the two clocks is actually costing you the mentions.
References
[1] Erin T. Jacques, Erela Datuowei, Elizabeth Quaye, Corey H. Basch, Arijit Chatterjee, Juanita Davis. "Authority Signals in Claude AI Health Citations: A Descriptive Analysis Using the Authority Signals Framework." arXiv, 2026 (preprint). https://arxiv.org/pdf/2605.23921
[2] Anthropic. "Enable and use web search." Anthropic Help Center. https://support.anthropic.com/en/articles/10684626-enabling-and-using-web-search
[3] Anthropic. "Claude web search now available globally on all plans." anthropic.com, March 20, 2025 (updated May 27, 2025). https://www.anthropic.com/news/web-search
[4] Anthropic. "Web search tool." Claude Platform Docs. https://docs.anthropic.com/en/docs/build-with-claude/tool-use/web-search-tool
[5] Anthropic. "Citations." Claude Platform Docs. https://docs.anthropic.com/en/docs/build-with-claude/citations