There is no "optimize for AI": each engine cites a different web
Marketers talk about "AI search" as if it were one thing to optimise for. It is not. ChatGPT, Gemini and Perplexity reach for different sources to answer the same question — and a page that one of them quotes constantly, another may never surface at all.
the short answer
AI search is not one engine. A 2025 analysis of 6.8 million citations found the engines pick different sources: Gemini favours your own structured site (about 52% of its citations were brand-owned), ChatGPT leans on directories and references, and Perplexity favours industry directories and community content. Overlap is low, so there is no single "optimise for AI." Build the shared foundation — clear entity, extractable structure, crawlability — then cover the distinct signal each engine rewards.
key takeaways
- "AI search" is not one system: a 2025 analysis of 6.8 million citations found ChatGPT, Gemini and Perplexity choose noticeably different sources.
- Gemini favours your own structured site — about 52% of its citations were brand-owned — rewarding schema and clean, factual pages.
- ChatGPT leans more on directories, listings and reference sources; Perplexity favours industry directories and community content.
- Overlap between engines is low, so a brand can dominate one and be absent from another for the same query.
- There is no single "optimize for AI": build the shared foundation, then cover the different source signals each engine rewards.
same question, different sources — 6.8M-citation analysis, 2025
In words, so the table does not carry it alone: an analysis of 6.8 million citations across the three engines found they do not source the same way. Gemini drew most of its citations from brands' own structured sites — around 52% were brand-owned — rewarding schema and clean factual pages. ChatGPT leaned more on directories, listings and reference sources than on any single brand's domain. Perplexity sourced more narrowly, favouring industry directories and community discussion. Three engines, three different ideas of what a trustworthy source looks like.
What the divergence actually means
The headline is easy to misread as "some engines are better than others." That is not it. The point is that each engine has its own taste in sources, formed by how it was built and what it has learned to trust, and those tastes are different enough that visibility does not transfer. You can be the source Gemini reaches for on a topic and be entirely absent from ChatGPT's answer to the same question — not because your page got worse between tabs, but because the two engines are looking in different places and weighting different signals when they decide whom to quote.
That breaks a comfortable assumption a lot of teams still carry over from search: that there is a single ranked truth, and being good lifts you everywhere at once. In classic search, optimise well and you tended to rise across the board because everyone queried roughly the same index. AI search fragmented that. There is no shared ranked list underneath the engines — each assembles its own answer from its own preferred sources — so "we did the AI optimisation" is no longer a single, checkable claim. It is three, and they can come back with three different verdicts.
The fragmentation, in three parts
Why the engines diverge, what each one rewards, and why that ends the dream of a single AI playbook. Open each layer for the part that changes how you work.
01 Why the engines diverge
Each engine was built on a different foundation and learned to trust different things. Gemini lives inside Google’s ecosystem and behaves like a search engine with strict sourcing, favouring structured content straight from a brand’s own domain. ChatGPT browses from a web index and its own model knowledge, leaning on reference sources and directories more than on any one brand’s pages. Perplexity sources more narrowly still, favouring vertical directories and community discussion. They are not three windows onto the same ranked list — they are reading overlapping but distinct slices of the web through different notions of what counts as a credible source. That is why the same page lands very differently across them.
02 What each one actually rewards
The 6.8-million-citation analysis makes the differences concrete. For Gemini, the lever is your own site: structured, factual, schema-rich pages and consistent data, because a majority of its citations came from brand-owned domains. For ChatGPT, the lever is presence in the references and directories it draws on, plus a clear entity its model already recognises — your own blog matters less. For Perplexity, the lever is showing up in the specific industry directories and community sources it leans into. Same brand, three different to-do lists: own-site structure for one, third-party presence for another, vertical directories for a third — and a page optimised for only one of them is leaving the other two on the table.
03 Why this kills the "one AI strategy" pitch
If the engines cited the same sources, a single playbook would cover them all. They do not, and the overlap between which domains each cites is low enough that a brand can be everywhere in one engine and nowhere in another for the same question. That breaks the tidy promise of "optimise for AI" as one task. It does not mean the work triples — much of the foundation is shared — but it does mean you cannot declare victory because you show up in the one engine you happened to check. The honest version of an AI-visibility strategy names the engines, measures each, and treats "we are cited" as a per-engine fact rather than a single status light.
Start with the entity, the one anchor they all check
If the shared foundation has a single most-valuable piece, it is your entity — the machine-readable answer to "who is this and what do they do." Engines that diverge on everything else still converge on needing to know that, and they confirm it against the same handful of references: a Wikidata record, a Wikipedia entry where one is warranted, consistent listings, and the same facts repeated across sources they trust. A brand whose name, description and core facts are consistent everywhere is easy for any model to recognise; one whose details conflict across the web makes every engine less sure who it is reading, and uncertainty is the enemy of citation.
This is why entity work pays across the fragmentation rather than into one engine. It is not a Gemini tactic or a ChatGPT tactic; it is the layer beneath all of them, the thing that lets a model connect a mention on one source to a page on another and decide they describe the same trustworthy thing. Get it wrong and the engine-specific work is built on sand — you can be in all the right directories and still be cited as someone else, or not recognised at all. Get it right and every later effort lands harder, because the model already knows who is being talked about.
How to turn this into a measurement, not a guess
Because the engines diverge, "are we visible in AI" is not a question you can answer once. The only honest way to know is to ask each engine the questions your buyers actually ask and record who it cites — engine by engine, prompt by prompt. A brand that checks only the engine its team happens to use walks away with a false read: cited there, it assumes it is cited everywhere, when the next engine over may never name it. Treat each engine as its own scoreboard, and the picture stops being a single reassuring impression and becomes a list of specific gaps you can act on.
The measurement also tells you where to spend, which the averages never will. If you are cited well in the engine that rewards owned content but absent from the ones leaning on directories and community, the gap names your next move — and it is not "more of what already works." Without per-engine data you optimise for the surface you can already see and quietly cede the others. With it, you stop guessing which signal is missing and start closing the specific one that keeps a given engine from naming you.
The mistakes the fragmentation punishes most
Three errors follow directly from forgetting that the engines differ. The first is declaring victory from one engine: a team sees its brand quoted in the tool it uses daily and assumes the job is done, never checking the others where it may be absent. The second is copying a competitor's single-engine playbook wholesale — if they earned their visibility in the engine that rewards owned content and your buyers live in the one that leans on communities, their tactics will not transfer, and you will have spent a quarter optimising for an audience that is not yours.
The third is subtler and more common: treating the shared foundation as optional because it is not tied to a specific engine. Entity consistency, extractable structure and crawlability feel like background work, so they get deferred in favour of whatever engine-specific tactic is trending. But skip them and every engine-specific effort underperforms, because the model cannot cleanly recognise or read what you publish. The fragmentation does not reward cleverness in one engine; it rewards doing the universal work first and the targeted work second, in that order, every time.
Engines and sources: quick answers
Do I need a different strategy for each AI engine?
You need one foundation and different emphases, not three unrelated strategies. The engines pull from overlapping but distinct source pools, so the same brand can be cited heavily by one and ignored by another. A 2025 analysis of 6.8 million citations found Gemini drew most of its citations from brands’ own structured sites, while ChatGPT leaned on directories, listings and reference sources, and Perplexity favoured industry directories and community content. The practical reading is not "build three sites" but "make sure you are strong on each kind of signal": a clean, structured, schema-rich site for the engines that reward owned content, accurate presence in the directories and listings others lean on, and earned mentions in the communities and references a third group trusts. One brand, several surfaces it has to be legible to.
Why does ChatGPT cite different sources than Gemini?
Because they were built on different foundations and trust different things. Gemini sits inside Google’s ecosystem and behaves more like a search engine with strict sourcing — it favours structured, factual content straight from a brand’s domain, rewarded with schema and consistent data. ChatGPT’s browsing starts from a web index and its own model knowledge, and leans more on reference sources and directories than on any single brand’s pages. They are not reading the same web through the same lens, so the set of sources each finds most quotable diverges. This is why a page perfectly tuned for one can be invisible in another: you optimised for one engine’s notion of a trustworthy source, not the other’s.
What works across all of them?
A few things travel well because every engine needs them. A clear, consistent entity — the same name, description and facts about you across the web, anchored where engines check, like Wikidata and reputable references — helps a model recognise you regardless of which sources it favours. Extractable structure helps too: a direct answer near the top, clean headings, short self-contained passages a model can lift cleanly. And being reachable matters everywhere — a page no crawler can read is cited by none of them. Past that shared core, the engines diverge, so the cross-platform play is to nail the universal signals first, then shore up whichever engine-specific source type you are weakest on.
Should I prioritize one engine?
Prioritise by where your buyers actually are, not by which engine is easiest. If your audience researches in ChatGPT, being strong in the directories and references it leans on matters more than perfecting a signal Gemini rewards; if they live in Gemini, your own structured site does more work. The mistake is spreading effort evenly across engines you have not checked your audience uses, or assuming the biggest engine is automatically your priority. Measure where you are cited and where your buyers ask, then weight the engine-specific work toward the overlap. The shared foundation — clear entity, extractable structure, crawlability — you build regardless; the prioritisation is about which engine-specific source pool you reinforce first.
A note on sources and certainty
The central figures — the 6.8 million citations analysed, Gemini's roughly 52% brand-owned share, the directory and community tilts of ChatGPT and Perplexity — come from a 2025 cross-engine citation study. We present them as a snapshot, not a constant: source preferences shift as engines retune, and the exact percentages will move, so treat the direction (the engines genuinely diverge) as the durable finding and the decimals as a moment in time. We have deliberately kept to what was measurable as of this writing and left later, sharper numbers for when they are real. What is unlikely to reverse is the core fact: there is no single AI index, and "optimised for AI" is a per-engine claim. The AC Group has spent 27 years treating each discovery surface on its own terms rather than assuming one rule fits all; this is that habit applied to a search layer that has quietly split into several.