Which AI search engine should you optimize for first?

Optimize first for the engine your buyers actually use, which is usually visible in your own referral data before it is visible in any market report. Check the referrer breakdown in your analytics and your server logs for chatgpt.com, perplexity.ai, gemini.google.com and copilot.microsoft.com, and work on whichever is already sending people, because that engine has demonstrated it can retrieve you.

When the data is too thin to read, which it often is at low volumes, size is the next best tiebreaker. According to Similarweb's AI tracker, ChatGPT held 64.6% of worldwide generative AI website traffic in January 2026 and Gemini 22%, so the two of them together account for most of the audience, and a first pass aimed at both covers more ground than a careful programme aimed at one small engine.

There is a second consideration that overrides both for some businesses. If your category is researched rather than casually asked about, Perplexity and the deep research modes matter more than their traffic share suggests, because they crawl more sources per answer and reach further down the ranking. A supplier chosen after forty minutes of research is worth more than a name recalled in a chat.

What you should not do is treat "AI" as one destination. The engines read different indexes, crawl differently and cite different pages, and a programme that assumes work on one transfers to the others will quietly underdeliver on three of the four.

How big is each AI engine in 2026?

The engines differ in size by more than an order of magnitude, and the gap has narrowed sharply in one year. Similarweb's global AI tracker reported ChatGPT at 64.6% of worldwide generative AI website traffic in January 2026, against 86.7% twelve months earlier, with Gemini rising from 5.7% to 22% over the same period.

Engine	Share of gen AI website traffic, Jan 2026	Same measure, Jan 2025	Why it matters for citation
ChatGPT	64.6%	86.7%	Largest single audience, and the one most buyers name
Gemini	22%	5.7%	Shares Google's crawling and rendering pipeline
Grok	3.5%	Not reported in the same set	Fast growth, little published citation research
DeepSeek	3.3%	Not reported in the same set	Largely outside Western buying journeys
Claude	2.1%	1.5%	Small audience, heavily weighted to technical buyers
Perplexity	1.9%	1.9%	Cites more sources per answer than its share suggests
Microsoft Copilot	1.1%	1.5%	Reaches people inside Microsoft 365 rather than on the web

Read the column for what it measures, which is visits to the assistants' own websites. Traffic share is not query share, and it misses assistants embedded inside other products, so Copilot in particular is undercounted by any measure based on visits to a domain. Google's AI Overviews sit outside the table entirely, because they appear inside ordinary search results rather than on an assistant site.

The trend is the part to plan against. One engine holding four fifths of the audience made a single-engine strategy defensible in 2025. Two engines splitting roughly six sevenths of it, with the second one attached to the world's largest crawler, makes the same strategy a gap.

Do the AI engines cite the same pages as each other?

The engines mostly do not cite the same pages, and the size of the difference is the single most useful number in this article. An analysis of 680 million citations reported by Averi in 2026 found roughly 11% of domains cited by both ChatGPT and Perplexity, and an independent Whitehat SEO study of 118,000 responses arrived at the same figure.

Overlap measured across other engine pairings lands in a similar band in the published datasets, from single digits to the low teens, and no study so far has found two engines drawing from broadly the same set of domains. Whichever dataset you prefer, the conclusion survives: being cited by one engine tells you very little about whether another will cite you.

Two things explain most of the gap. The engines read different indexes, so the candidate set differs before any ranking happens, and they differ in how many sources they pull per answer, with Perplexity citing substantially more sources per response than ChatGPT. An engine that cites ten sources reaches further down its ranking than one that cites three.

The practical consequence is that measurement has to be per engine and so does the work. A report that says "we are cited in AI search" without naming the engine and the prompt is not a measurement. Our guide to tracking AI visibility across engines sets out what to record from each answer, and why ChatGPT cites you when Google does not works through the most common version of the split.

Which index does each AI engine actually read?

Each engine reads a different index, and knowing which one explains most of the differences in citation behaviour. Google feeds AI Overviews and AI Mode from its own crawl, which is why a page Googlebot has rendered can appear there while being absent elsewhere, and why Gemini is the engine most closely tied to classic SEO fundamentals.

ChatGPT has been built on a Bing partnership for web results while OpenAI has been crawling with OAI-SearchBot to build an index of its own, so both being present in Bing and being fetchable by OpenAI's crawler matter. Microsoft Copilot sits directly on Bing, which means Bing Webmaster Tools is the closest thing to a console either of them offers you.

Perplexity runs its own crawler, PerplexityBot, its own index and its own ranking, having moved off third-party search infrastructure as it grew. Anthropic's Claude has been reported to use the Brave Search API for its web results, which is a different index again, and one with its own crawler and its own coverage gaps.

None of these arrangements is permanent, and none is documented by the engines in the detail an SEO would want, so treat the mapping as current best understanding rather than as specification. What follows from it is stable even as the details change: a page has to be crawlable by several independent crawlers to be eligible everywhere, and our page on whether to block AI crawlers covers which user agents to allow.

How many searches actually show an AI answer?

The share of searches that show an AI answer is contested, and the honest range is wide. Semrush's study of more than 10 million keywords tracked AI Overviews appearing on 6.49% of queries in January 2025, rising to 24.61% in July and settling near 15.69% by November 2025. Conductor's benchmark of 21.9 million queries found 25.11% in the first quarter of 2026, and BrightEdge's commercial tracker has reported figures closer to 48%.

The spread comes from the query sets rather than from bad measurement. A panel weighted toward informational keywords produces a high number, a mixed or transactional set produces a low one, and Google has continued to expand coverage into commercial queries where it used to show almost none. Anyone quoting a single percentage without saying which keyword set produced it is quoting a coincidence.

For planning purposes, the defensible statement is that a substantial minority of searches in most categories now return a generated answer, and the share is rising rather than falling. That is enough to act on without pretending to a precision the published data does not support.

What matters more than the headline share is your own category. Run thirty of your real buyer questions through Google and count how many return an AI Overview, and you will have a number that applies to your market rather than to an index panel. Our note on whether AI search sends traffic covers what happens to clicks when one appears.

How do you find out which engines your own buyers use?

Finding out which engines your buyers use takes three checks, none of which needs a tool. Start with referral traffic: in Google Analytics 4, look at the session source and medium report for chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com and the various Bing referrers, and treat any of them appearing at all as evidence that engine can already retrieve you.

Next read your server logs for the crawlers rather than the humans. GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot, Bingbot and Googlebot each tell you something different: which of them fetch your pages, how often, and which URLs they take. An engine whose crawler never arrives is not going to cite you, and that is a fixable access problem rather than a content problem.

Then ask the people on calls. Sales teams hear "I asked ChatGPT and it suggested you" or "I found you through Perplexity" long before the pattern is visible in aggregate data, and one week of asking every inbound enquiry where they looked is worth more than a quarter of dashboard watching at low volume.

Run all three and a priority usually announces itself. Where it does not, sample the same prompt set across all four engines on a schedule and let the citation counts decide, which is the approach our methodology takes, and record the prompt, engine, date and every cited URL so the comparison holds up over time rather than reflecting one lucky run.

What else do people ask about choosing an AI engine to optimize for?