Claude and ChatGPT Recommend Different Businesses. Here Is Where Each One Looks

By Matt Griffin, Founder, Formative Digital · Published 18 July 2026

That table is the whole argument of this page compressed into five rows. Everything below unpacks why the two engines diverge, what our own testing found when we asked them the same questions, and what a business in Ontario should actually do about it. This piece maps to Vector 1, Diagnose, and Vector 7, Distribute, in our 12-vector process: first establish where each engine looks, then earn a presence in exactly those places.

Why "Claude vs ChatGPT" is a visibility question, not a product review

Most comparisons of these two assistants argue about which writes better code or nicer emails. That debate matters to people choosing a subscription. It does not matter to a business owner, because your customers already chose for you: some of them ask ChatGPT for a recommendation, some ask Claude, and both groups act on the answer without ever opening a results page.

Framed that way, the two products stop being rival chatbots and become two separate discovery channels, the way Google and Bing were two channels in 2010, except the gap between them is wider. Google and Bing at least crawled the same open web and converged on similar answers. Claude and ChatGPT retrieve from different indexes, license data from different partners, and were trained by companies with visibly different philosophies about what a trustworthy answer looks like. The result is two referral streams with different rules, and a business that only satisfies one set of rules quietly forfeits the other stream.

How Claude decides which sources to trust

Claude's live web search is powered by the Brave Search index. This is documented behaviour, not speculation: a 2025 analysis by Profound, reported by Search Engine Land, found that 86.7% of the sources Claude cited matched Brave's top organic results for the same query. Claude does very little re-ranking of what Brave hands it. If Brave puts you on the first page, Claude can quote you; if Brave has never indexed you, Claude cannot, and your Google positions are irrelevant to the outcome.

The same research showed which page shapes Claude reaches for. A majority of its citations came from blog-style paths, nearly half from comparison and list formats, and about a quarter of cited URLs carried a year token. Claude also triggers search most aggressively on "best", "top", and versus-style prompts. Put plainly: Claude behaves like a careful researcher who pulls the top Brave results for a comparison query and reads them before answering, which is why well-structured comparison content earns a disproportionate share of its citations.

Our own engine testing adds a second layer. When we ran identical business-recommendation prompts across four AI engines for the consensus gap study, Claude leaned noticeably on curated directory-style sources: professional associations, vetted listing sites, established review platforms. Where ChatGPT would happily name a brand it simply recognised, Claude preferred a name it could tie back to a source it had just retrieved. That caution is a feature for users and a filter for businesses: to appear in Claude answers you need to exist in the curated layer of the web, not just on your own domain.

The training difference behind the caution

Claude's sourcing behaviour is not an accident of engineering; it traces back to how Anthropic trains the model. The method, described in the 2022 Constitutional AI paper by Bai and colleagues, replaces most human labelling of harmful outputs with a written list of principles the model uses to critique and revise its own answers. The goal stated in the paper is "a harmless but non-evasive AI assistant", one that engages with hard questions while explaining its reasoning instead of dodging or bluffing.

Two consequences of that design show up in search visibility. First, a model trained to critique its own drafts against explicit principles carries that self-checking habit into answer generation, which is consistent with the tight coupling we observe between Claude's claims and its retrieved citations. Second, the Anthropic authors are candid that every deployed system embodies some value choices, writing that "we cannot avoid choosing some set of principles to govern it". Claude's principles push toward transparency and attributable claims, so the businesses it names tend to be ones with an attributable paper trail: consistent entity data, third-party validation, sources worth citing.

ChatGPT's lineage runs through reinforcement learning from human feedback, tuned across an enormous consumer base, and its answers read accordingly: fast, confident, and weighted toward brands with heavy footprints in its training data. Neither approach is wrong. But they reward different evidence, and your visibility work should supply both kinds.

How ChatGPT builds a business recommendation

ChatGPT's discovery stack has two distinct layers. For general questions it retrieves from the Bing index, which means classic Bing signals, indexation, IndexNow, Bing Webmaster Tools health, still buy relevance in 2026. For location-based questions it does something most owners have never heard of: it fires an internal tool call to licensed data partners, and since a December 2024 partnership the dominant partner is Foursquare. Multiple 2026 analyses put Foursquare's share of the business names ChatGPT surfaces for local queries at 60% or higher, with Bing Places filling gaps.

Sit with that for a second. A directory many Canadian businesses abandoned a decade ago now feeds the recommendation engine used by hundreds of millions of people. In audits of Ontario service businesses this year, the Foursquare check is the one that most reliably embarrasses everyone in the room, ourselves included the first time we ran it: claimed Google profiles, active review management, and a Foursquare listing showing an old address and a category from 2014.

On top of retrieval, ChatGPT carries strong prior knowledge of brands from training. Widely mentioned companies get named even without a fresh retrieval supporting the mention. That favours incumbents and franchises, the Goliaths, and it is precisely the layer where an independent business needs deliberate work: press mentions, industry publications, and community references that put the brand name into the corpus. The step-by-step version of that work lives in our guide to optimising a website for ChatGPT.

What happens when you ask both the same question

The divergence is measurable. In our consensus gap testing we put identical recommendation prompts to ChatGPT, Claude, Gemini, and Perplexity and compared the shortlists. The overlap was thin enough to surprise us: for many Ontario service categories, the business named first by one engine did not appear at all in another engine's answer. Different indexes, different partners, different training priors, different shortlists.

Matt's read on what this means for owners: "The mistake I keep seeing is a business owner asking one engine about themselves, liking the answer, and closing the tab. Claude and ChatGPT are not two doors into the same room. They are different rooms, and I have watched a client dominate one while being completely absent from the other until we fixed the specific source that engine reads." Full methodology and category-level results are in the consensus gap research; the honest summary is that single-engine visibility checks tell you almost nothing about the other engines.

What Claude rewards: the curated-layer checklist

Claude visibility work concentrates on getting into the sources Claude retrieves and trusts. The specific levers, in the order we apply them for clients:

• Brave indexation first. Confirm your key pages appear in Brave Search for your money queries. This is the gating condition; nothing else on this list matters until it passes.
• Publish comparison and evaluation content. Claude's citation profile over-selects versus pages, list formats, and dated posts. A truthful "X vs Y for [your market]" page on your domain is a direct bid for its citations.
• Get into the curated directories for your profession. Association memberships, accreditation bodies, vetted regional listings. These are the sources Claude reached for repeatedly in our testing.
• Keep claims attributable. Pages that state numbers with named sources survive Claude's cautious synthesis; pages full of unsupported superlatives get skipped.
• Date your content honestly and refresh it. Year tokens appeared in roughly a quarter of Claude's cited URLs. Freshness is a retrieval signal here, not a vanity metric.

The extended version of this list, including how Brave crawls differ from Googlebot, is in our dedicated guide to Claude optimisation.

What ChatGPT rewards: entity weight and partner data

ChatGPT work splits along its two layers. For the retrieval layer: be indexable and healthy in Bing, submit through IndexNow, and structure pages so a retrieved chunk answers the question on its own. For the partner layer: claim and correct the Foursquare listing, verify Bing Places, and make the name, address, phone, and category identical everywhere the engines might look. Directory consistency sounds like 2015 advice; the Foursquare partnership made it 2026 advice again.

Then there is the training-data layer, the slowest and most valuable. Every legitimate mention of your brand in industry press, local media, supplier pages, and community coverage raises the odds that ChatGPT names you from memory when no retrieval fires. This is Vector 7, Distribute, and it is the part of the work that most resembles old-fashioned PR. For an independent business the encouraging news is that AI engines weight relevance and consistency more mechanically than human journalists ever weighted fame, so a well-documented local specialist can outrank a vaguely-known national brand inside its own category.

The markup both engines can read

Structured data is the one investment that pays on both surfaces at once, because every retrieval pipeline parses schema to disambiguate entities. At minimum, a local service business should ship a LocalBusiness node whose details match its directory listings exactly. Here is the shape we deploy, trimmed to essentials:

{
  "@context": "https://schema.org",
  "@type": "LocalBusiness",
  "name": "Example Plumbing Co.",
  "address": {
    "@type": "PostalAddress",
    "addressLocality": "Brantford",
    "addressRegion": "ON",
    "addressCountry": "CA"
  },
  "telephone": "+15195550123",
  "url": "https://exampleplumbing.ca",
  "sameAs": [
    "https://foursquare.com/v/example-plumbing",
    "https://www.bing.com/maps?ss=ypid.YN1234"
  ]
}

The detail that does the disambiguation work is sameAs: it explicitly ties your web entity to the directory records ChatGPT's partners hold and the listing pages Brave serves to Claude. When the graph on your site and the data in the directories agree, both engines can connect the pieces with confidence; when they disagree, you become the ambiguous entity neither engine wants to recommend. Schema design of this kind sits inside our GEO service, and it is Vector 6, Structure, doing double duty across engines.

The both-engines playbook, sequenced

Roughly two thirds of the work overlaps between engines, which is what makes a combined program affordable. The sequence we run:

1. Diagnose on both surfaces. Ask each engine the questions your customers ask, and record who gets named and which sources appear.
2. Fix the entity layer. Identical NAP and category data across your site, Foursquare, Bing Places, Google Business Profile, and your professional directories.
3. Verify both indexes. Brave for Claude, Bing for ChatGPT. Neither engine can cite a page its index lacks.
4. Ship structured data. LocalBusiness or Organization with sameAs links to every directory record.
5. Publish citable comparison content. This targets Claude's citation habits directly and feeds ChatGPT retrieval at the same time.
6. Earn third-party mentions. Press, associations, and community coverage build the brand weight ChatGPT draws on and the curated presence Claude checks.
7. Re-test quarterly. Partner deals change; the December 2024 Foursquare agreement rewired ChatGPT local answers overnight, and the next such deal will rewire something else.

A necessary honesty note, since visibility claims in this space run hot: nobody controls what a generative model says, including us. The seven steps above change what each engine can find and verify about you, which in our testing changes what it says far more often than not. That is influence built on evidence, and evidence is the only lever that survives model updates.

Which engine deserves your first dollar

If your customers decide quickly and locally, restaurants, trades, retail, start with the ChatGPT side: the audience is many times larger, and the fixes (Foursquare, Bing Places, entity consistency) are cheap and fast. If you sell considered services, law, accounting, engineering, B2B, weight Claude earlier: its user base skews toward professionals doing careful evaluation, exactly the buyer who compares three firms before calling one. And if you are choosing between engines because the budget only covers one, revisit the overlap: steps two through five of the playbook serve both, so the real decision is only about where the engine-specific effort goes first.

What we would not do is wait. Every quarter a business stays absent from these answers, the engines keep recommending whoever is present, and those recommendations compound into reviews, mentions, and stronger presence for the competitor. The gap widens on its own.

Frequently asked questions