Essay · 09

What ChatGPT, Perplexity, Claude and Google's AI answers each reward, and where they diverge

14 July 2026 · 7 min read

In June I ran the same set of buyer questions, things like "best AI SDR tool for a ten-person sales team", through ChatGPT, Perplexity, Claude and Google's AI answers on the same day, as part of a measurement I published here. The four engines gave me four different reading lists. Perplexity cited community threads and YouTube reviews. Google's AI Overview leaned on pages that already ranked, plus Reddit. ChatGPT and Claude each quoted a handful of tidy explainer pages, and not the same handful. Exactly one vendor appeared in all four answers.

That result is the whole argument of this essay. "AI search" gets discussed as if it were one channel with one playbook. It's four engines with four editorial tastes, and if you plan for the average of them you're planning for an engine that doesn't exist. Here's what each one rewards, what they all reward before anything diverges, and, the part most write-ups skip, how much of your plan should actually change per engine. The honest answer is: less than the differences suggest.

What all four reward before anything diverges

Roughly four fifths of the work is identical across every engine, and it has to come first. All four need to be able to read your site, which is less of a given than it sounds: since July 2025 Cloudflare has blocked AI crawlers by default on new sites, and plenty of owners are invisible without knowing it. All four prefer pages that answer one question cleanly, with the answer up front, over pages that circle a topic for a thousand words. And all four lean on third-party corroboration: Muck Rack's Generative Pulse study put earned media at around 84% of AI citations (May 2026), and my own scan logs agree with the direction of that number, if not always the size. A brand that only publishes about itself, on its own site, reads as a claim. The same facts repeated by sources the engine already trusts read as a fact.

If that foundation is missing, per-engine tactics are decoration. Every divergence below assumes it's in place.

Where each engine's taste diverges

Perplexity: the generous one

Perplexity attaches numbered citations to every answer, so it cites more sources per question than any of the others. In my June scan of the AI SDR category, the most-cited vendor earned 30 top-three citations, and 28 of them were Perplexity's; ChatGPT supplied the other two. That ratio held down the leaderboard. Its taste runs to community discussion and video: Reddit threads, review sites, YouTube walkthroughs show up in its citation lists far more often than in ChatGPT's or Claude's. For most brands Perplexity is the early-warning system. If you can't earn a citation there, with the lowest bar of the four, the problem is your content or your corroboration, not the engine.

Google's AI answers: the incumbent's shortcut

AI Overviews are assembled largely from pages that already rank in the underlying search results, so a decade of ordinary SEO carries over more directly here than anywhere else. The second ingredient is community content. Google signed a licensing deal with Reddit in early 2024, reported at around $60m a year, and the citation patterns since have matched the receipt: Reddit threads appear in AI Overviews at a rate no other engine matches. If your category gets discussed on Reddit and you're absent from those threads, this is the engine where it costs you most.

ChatGPT: the biggest audience, the stingiest citations

ChatGPT has the largest share of buyer conversations, and it cites the least generously. Its search draws on Bing's index plus OpenAI's own crawling, and answers typically quote a small number of sources, favouring self-contained explainers: pages that define the thing, answer the question, and stop. There's also a wrinkle the other engines don't have to the same degree: ChatGPT answers plenty of questions from what its model already knows, with no live search and no citations at all. For those answers, being well represented in widely-crawled third-party sources over years is what puts you in the answer. You can't retrofit that in a quarter, which is an argument for starting now rather than a reason to despair.

Claude: the conservative one

Claude added live web search in 2025 and uses it the way a careful researcher would: fewer sources, weighted toward documentation-style pages, and a visible reluctance to cite anything that reads as marketing. In my logs it's the stingiest citer of the four. The practical consequence: plain, structured, factual pages, the kind that state what a thing is, what it costs and what its limits are, outperform persuasive copy here by a wider margin than anywhere else. Claude's traffic share is the smallest of the four for most categories, but its users skew technical, which for some ICPs makes it the highest-value citation on the board.

Where the divergence should actually change your plan

Three places, and only three, in my experience so far.

First, distribution. The shared foundation says "earn third-party corroboration"; the divergence says where. If your buyers lean on Google and your category lives on Reddit, community presence outranks press. If they're Perplexity-type researchers, YouTube and review sites move the needle. Same principle, different target list, and your measurement log, which records who gets cited instead of you per engine, is what makes the choice for you.

Second, prioritisation. You don't have four engines' worth of effort, so weight by where your buyers actually are, which you learn from referral analytics and from asking customers, not from market-share headlines. A developer-tools founder should care about Claude citations far more than a DTC brand should.

Third, measurement. Because the engines disagree, an averaged citation rate hides exactly the finding you need. "Cited in Perplexity, invisible in ChatGPT" is a diagnosis: it usually means the content is fine and the corroboration is thin. I covered the method in how to measure AI citations; the short version is to keep the four numbers separate, always.

What I got wrong about this

Two admissions. When I started measuring last year I treated the engines as one channel and planned to report one citation rate. That lasted one scan cycle: the per-engine numbers diverged so much that the average was meaningless, and I rebuilt the log to keep them apart. I also assumed Google's AI Overviews could be safely ignored as repackaged SEO, on the logic that anyone doing SEO properly was covered by default. Wrong in one important way: the Reddit weighting means a brand can rank respectably in classic results and still be absent from the AI answer, because the answer is drawing on community threads the brand was never part of. The overlap with SEO is real, but it isn't total, and the gap is precisely where the opportunity sits.

There's a trade-off I should name too. Chasing each engine's taste has a cost, and the tastes shift; any specific weighting I've described here will drift as the engines re-tune. The shared foundation doesn't drift. When in doubt, put the marginal hour there.

What changed in the work

Three concrete changes. My scan logs record citations per engine, never averaged, and the "cited instead of you" list is kept per engine too, because it's effectively four different target lists. Audit reports now lead with an engine-weighted view: which engines this client's buyers actually use, and how the citation gap looks in those specifically. And my own content, including this site, is written to the shared foundation first, one question per page, answer up front, corroboration earned off-site, with engine-specific distribution decided by the measurement rather than by hunch. Case Study Zero runs exactly this way, in public: the baseline is cited in zero of seven buyer queries, logged per engine, and if the pattern in this essay holds, the first citation to land should be a Perplexity one. That's a prediction you can check me on.

Common questions

← All essays