AI search optimization: five engines, mostly different sources

79.61% of the websites that AI engines cited in the first half of 2026 showed up on one engine only. That figure comes from the cross-engine citation study I ran on Wellows' citation data: 13,037,251 citations across 531,889 questions that all five engines answered, from January to June 2026.

So "AI search" isn't one place you optimize for. It's five retrieval systems that mostly reach for different sources, and four in five of the websites any one of them cites never show up on the other four. That's the whole argument of this page, and everything below is about what to do with it.

What AI search optimization is

AI search optimization is the practice of making a brand's pages the sources AI engines retrieve, quote and cite when they answer a question. The engines that matter today are ChatGPT, Gemini, Perplexity, Google AI Overviews and Google AI Mode. The goal is a citation, a visible link to your page inside the answer, not just a mention of your name.

The field has collected several names, and people searching for it use all of them:

Name	What it usually points at
AI search optimization	The umbrella term: every AI answer surface, measured per engine
Generative engine optimization (GEO)	Being cited inside answers that a model generates (ChatGPT, Gemini, Perplexity)
Answer engine optimization (AEO)	Being the quoted answer: featured snippets, People Also Ask, AI Overviews
AI SEO	Used two ways: SEO for AI engines, or SEO done with AI tools. Check which one someone means
LLM optimization (LLMO)	Same territory as GEO, framed around the language model
AI visibility	The outcome, not the practice: how often a brand is mentioned or cited across AI answers

I treat SEO, AEO and GEO as layers built in that order, not rivals. The full case is in SEO vs AEO vs GEO. Google's own guide to optimizing for generative AI features has a section on exactly this: applying foundational SEO best practices. It goes further than I do. From Google Search's perspective, the guide says, optimizing for generative AI search "is optimizing for the search experience, and thus still SEO." That's a statement about Google Search, and it covers two of the five engines on this page. The layers are how I handle the other three.

Five engines, mostly different sources

The engines don't just disagree a little. Across the study, the distribution of how many engines cited the same website looked like this:

Engines citing the same website	Share of cited websites
1 engine only	79.61%
2 engines	14.47%
3 engines	4.36%
4 engines	1.25%
All 5 engines	0.31%

Fewer than one website in three hundred was cited by all five. If you report "AI visibility" as one number, you're averaging five different results into a figure that describes none of them.

Each engine also leaves a different number of seats at the table. The mean number of distinct websites per answer ran Perplexity 4.68, AI Overviews 4.47, ChatGPT 4.15, AI Mode 3.97 and Gemini 3.77. Gemini gives you roughly one fewer slot per answer than Perplexity does, which matters when you're the fifth-best source on a topic.

Part of the reason may be plumbing. The engines don't all search the same index. Semrush's AI search guide (Carlos Silva, May 2026) describes AI Overviews retrieving from Google's search index, ChatGPT pulling web results through its own crawler (OAI-SearchBot) and Bing's index, and Perplexity using its own index. Different indexes feeding different answers fits a world where four in five cited websites are single-engine, though my study doesn't test the cause.

My single-engine number is lower than others published. Kevin Indig's Growth Memo (18 May 2026) landed at 91% across three engines. Fewer engines in the denominator push that share up, and definitions of what counts as "a website" move it too: two alternative domain-normalisation rules put my own figure at 85.1% and 93.2%. Every version agrees on the direction. Most cited sources are single-engine.

Two findings I published, then had to take back

The study went live on 29 July 2026. Five days later I'd revised two of its headline findings, and both revisions change what you should optimize. The full record is in I tried to break my own AI study.

The first one looked like it settled the brand debate. Raw agreement between engines ran 30.3% at brand level against 6.76% at page level, so brand seemed 4.5 times more durable than any single page. Then I ran a permutation null: pair one engine's sources for a question with another engine's sources for a random different question, and see how much agreement you get by chance. Against that baseline, page agreement is 42 times chance, brand 17 times, website 11 times. The ladder reverses.

Brand is still the easier thing to hold across engines, because an answer only has a handful of plausible brands to choose from. But when two engines independently land on the same specific page, that's the stronger signal. Pages are where the work compounds.

The second one was about Google. AI Overviews and AI Mode agreed on sources 24.5% of the time against 7.3% for cross-company pairs, which looked like proof that Google's two surfaces share a pipeline. Chance-adjusted, the premium collapses to 1.06 times. The two highest chance-adjusted pairs in the whole set are cross-company: Gemini and Perplexity at 19.7 times, ChatGPT and Perplexity at 19.1 times. Optimizing for "Google's AI" is still two jobs.

If you only read one section of this page, this is the one I'd keep. Both wrong findings were the comfortable ones.

Ranking still matters. It just isn't where the job ends

Nothing above says to stop doing SEO. An engine can only cite a page its crawler could fetch and read, which is why the LLM technical SEO checklist starts with crawler access and ends with citation.

What changed is that a ranking no longer guarantees the visit. On a publisher I worked with through a core update recovery, average position came back from 32 to 8.2 over twelve months in Search Console, and click-through rate at that position sat around 0.7%, far below what position 8 normally returns. My reading is AI Overview absorption: the answer gets served on the results page and the click never happens. Search Console didn't break out AI features during that window (twelve months to May 2026), so treat it as an inference. Google announced Generative AI performance reports for Search Console on 3 June 2026, after that window closed. The core update recovery case study has the numbers.

The pattern isn't unique to that site. Pew Research Center (July 2025) found that Google users who saw an AI summary clicked a traditional result in 8% of visits, against 15% for those who didn't.

So the recovery wasn't finished at rank. It needed the answer layer on top.

AI Overviews and AI Mode: one index, two jobs

Google has written down what its AI features require, so start there. Its guide says a page must be indexed and eligible to appear in Google Search with a snippet, and the site must be included in Search generative AI features in Search Console. Meeting all of that still doesn't guarantee the page gets crawled, indexed or served. Google says so in the same paragraph.

The guide names two techniques behind both surfaces. Retrieval-augmented generation, which Google also calls grounding, uses the core ranking systems to pull relevant pages from the Search index, then shows clickable links to the pages that support the answer. Query fan-out is a set of related queries the model runs alongside yours: Google's example turns "how to fix a lawn that's full of weeds" into searches like "best herbicides for lawns" and "how to prevent weeds in lawn". Fan-out is why a page can be cited for a question it never targeted, and the Query Fan-Out Explorer is where I draft the sub-questions a page should answer before I write it.

A shared index doesn't mean shared sources. That was the second finding I had to take back: chance-adjusted, AI Overviews and AI Mode agree at 1.06 times the cross-company baseline, which is barely a premium at all. They also leave different room, 4.47 distinct websites per AI Overviews answer against 3.97 in AI Mode. Google's own AI features documentation says they "may use different models and techniques, so the set of responses and links they show will vary." Whatever the mechanism, I check them as two engines, because in the data they behave like two.

How to optimize for AI search, in the order I do it

The order matters more than the tactics, because each stage depends on the one before it.

1. Reach. Make sure the search and user-fetch bots (OAI-SearchBot, ChatGPT-User, Claude-SearchBot, PerplexityBot and the rest) can fetch the page and read the copy in the raw HTML. This site generates every machine-read layer from the page a person reads, which I wrote up in Built for Two Readers. For Google's two surfaces, add the eligibility rules above: indexed, snippet-eligible and included in generative AI features in Search Console. A page the crawler can't reach can't be cited, whatever else you do. Stages 2 and 5 of the checklist cover this.
2. Liftable passages. Write answers a model can quote without the paragraph around them: the answer in the first sentence under the question, numbers attached to their source. How to get cited by ChatGPT, Gemini and Perplexity walks through it, and the Citeability Checker scores a page against it. This isn't chunking. Google's guide says there's no requirement to break content into tiny pieces for AI, and nothing in this stage asks you to. It asks for the answer first and the source next to the number, which reads better for people too. The R2A Content Framework scores the rest of the path, from retrieval to action.
3. An unmistakable entity. Engines decide whose words to repeat partly on whether they can tell who you are. One central entity, named the same way across your site and your profiles, with the pages about it built as a topic cluster architecture that links back to one pillar.
4. Per-engine measurement. Track citations engine by engine, and segment LLM referral sessions in analytics. At Wellows, LLM referral sessions went from a 176-a-month baseline to roughly 9 times that over eight months, July 2025 to February 2026, in GA4. That's in the Wellows AI visibility case study. For what models remember about a brand, as opposed to what they retrieve, I use the BERAP Map.

Stage 4 is the one most teams skip, and it's the only one that tells you whether stages 1 to 3 did anything.

What Google says you can skip, and who it's speaking for

Google's guide has a section called "Mythbusting generative AI search". It lists five things you can ignore for Google Search:

Tactic	What Google's guide says
llms.txt and other "special" AI files or markup	Google Search doesn't use them. Keeping one "will neither harm nor help" your visibility in Google Search
"Chunking" content	No requirement to break content into tiny pieces, and "there's no ideal page length"
Rewriting content just for AI	AI systems understand synonyms, so you don't need every long-tail variation
Seeking inauthentic mentions	Core ranking systems focus on quality content and other systems block spam. The AI features depend on both
Overfocusing on structured data	Not required, and there's no special schema.org markup. Still worth using for rich results

The same guide warns off third-party tools that promise ranking success or claim internal Google metrics: "No third-party tool has access to our internal ranking or AI systems."

Not every page Perplexity cited for this keyword agrees. HubSpot's post on ranking in AI search calls llms.txt "a quick win" and suggests adding one. HubSpot's own tools guide, cited in the same answer, points to an Ahrefs study of 137,210 domains (Louise Linehan, June 2026) in which 97% of llms.txt files received no requests at all in May 2026. Same publisher, opposite advice, one engine citing both on the same day.

Google's list deserves a scope note. It speaks for Google Search, which is two of the five engines in my study, and the guide itself says it's fine to keep llms.txt "for other services or systems that use these files." So the accurate reading is narrower than "AI ignores llms.txt". Google's AI features ignore it. On this site I keep one as hygiene and never count it as a citation lever, which is where the checklist ranks it too.

Tools that measure AI search, and what each one can see

Most tool roundups sort by price. I'd sort by whose data it is, because a tool that blends five engines into one score has the problem from the second section of this page built in.

Tool	Whose data	What it can see
Search Console, Generative AI performance report	Google, first-party	Impressions in generative AI features on Google Search and Discover. Google is rolling it out to a subset of sites first
Bing Webmaster Tools, AI Performance	Microsoft, first-party	Citations in AI-generated answers, cited pages and grounding queries. Public preview since 10 Feb 2026
GA4 referral segments	Yours	Clicked citations only. ChatGPT links arrive tagged utm_source=chatgpt.com; an unlinked mention never becomes a session
Third-party monitors (Ahrefs Brand Radar, Profound, Peec AI, Otterly.AI and others)	The vendor's prompts	Observed answers and citations across engines, with coverage that depends on the plan. HubSpot's tools guide compares them
A dated prompt log	Yours	What each engine cited for one question on one day, like the table below

Neither first-party report covers ChatGPT or Perplexity. For those two you're relying on a monitor or a fixed set of questions you rerun, and the HubSpot guide makes the same point. Before trusting any monitor, I check two things: does it break results out per engine, and does it show the cited pages rather than just whether the brand appeared? A single visibility score is five different results averaged into one.

For a small site, a monitor may be more than you need. I logged the table below by asking each engine myself.

What ChatGPT and Perplexity cite for this page's own keyword

A six-month study tells you the shape. A single question tells you what it looks like on the day. On 6 October 2026 I asked ChatGPT and Perplexity the same question, "what is AI search optimization", and logged every source each one cited.

Engine	Date checked	Sources cited
ChatGPT	6 Oct 2026	seo2aio.ai, castle.co.uk, radiusrank.com, besalient.io (4 sources, all glossary or terminology pages)
Perplexity	6 Oct 2026	business.adobe.com, blog.hubspot.com (3 different posts), semrush.com, networksolutions.com, aleydasolis.com, marketingaid.io, developers.google.com (9 URLs from 7 domains)

ChatGPT and Perplexity cited eleven different domains between them and shared none. ChatGPT treated the question as a definition and cited only glossary pages from specialist AI-search sites. Perplexity went to guides and checklists from large publishers and SEO platforms, HubSpot three times over, plus Google's own guide. Every ChatGPT link also arrived tagged ?utm_source=chatgpt.com, which is the tag that lets GA4 attribute those visits to ChatGPT at all.

Gemini, AI Overviews and AI Mode join this table on the next rerun. I rerun it each quarter, so the date column tells you how fresh it is.

Where these numbers stop

The study is six months of data, January to June 2026. 84.06% of its questions were collected in the United States, across 27 markets in total, so it says less about engines in other languages. The permutation draws its comparison question from anywhere in the set, and a same-category null would put every multiple lower. And the finding that engines are converging (overlap rose 32.57% across the six months) has never been tested against a null, so I don't lean on it here.

None of that moves the 79.6%. It does mean your own category might sit somewhere else on the distribution, which is the reason to measure it rather than borrow mine.

Questions people ask about AI search optimization

What is AI search optimization called?

AI search optimization goes by several names: generative engine optimization (GEO), answer engine optimization (AEO), AI SEO and LLM optimization (LLMO). AI search optimization is the umbrella term. GEO and AEO describe two of its layers, being cited in generated answers and being the quoted answer on the results page.

Does AI search optimization replace SEO?

No. AI search optimization sits on top of SEO. An AI engine can only cite a page its crawler could fetch and read, so crawl access, indexing and clear structure still come first. What changes is the finish line: a citation inside the answer, measured per engine, instead of a ranking alone.

Is SEO still worth it in 2026?

Yes, because crawl access and clear, indexable pages are what every AI engine needs before it can cite you. The return on a ranking has changed, though. On one publisher, position 8.2 returned around 0.7% click-through, which is why ranking work now needs the answer layer added on top.

How do I optimize for Google AI Overviews?

Start with eligibility. Google's guide says a page must be indexed and eligible to show in Search with a snippet, and the site must be included in generative AI features in Search Console. After that it's foundational SEO and content that isn't a rewrite of what already ranks. Check AI Mode separately: in my study the two surfaces agreed on sources at only 1.06 times the cross-company baseline.

Do I need llms.txt or schema markup for AI search?

Not for Google. Its guide says Google Search doesn't use llms.txt and that no special schema is needed for its AI features, though structured data still helps with rich results. That guidance covers Google Search only, so I treat llms.txt as cheap hygiene and never as a citation lever. The Schema JSON-LD Generator writes the common structured data types.

How do I know if AI engines are citing my site?

Use the first-party reports where they exist: Search Console's Generative AI performance report for Google, and Bing Webmaster Tools' AI Performance report for Microsoft. For ChatGPT and Perplexity, track citations with a monitor or a fixed set of questions you rerun, and segment AI referral sessions in GA4. Check each engine on its own, because 79.61% of cited websites in my study appeared on one engine only.

Why does my brand appear in ChatGPT but not in Perplexity?

Because the engines mostly cite different websites. In my study, 79.61% of cited websites appeared on one engine only, and Perplexity shows more sources per answer (4.68) than ChatGPT (4.15). Visibility on one engine says little about the others, so check each one separately.

If you'd rather have someone run this for your site, the AI-search access check covers whether the right bots can reach and cite your pages.