Key takeaways
- Every AI answer combines what the model learned in training with what it retrieves live, then synthesises with citations.
- You can't edit training data, but you can shape retrieval (ranking), trust (entity and sources) and extractability (structure).
- Perplexity is retrieval-first and citation-transparent; ChatGPT blends training and retrieval; Gemini and AI Overviews lean on Google's index and Knowledge Graph.
- The highest-leverage work is the same across all of them: consistent entity, specific answerable content, presence on trusted sources.
The three layers of an AI answer
Ask an assistant “which agency should I use for technical SEO?” and three things happen in sequence.
Training. The model has read a large portion of the public web as of its training date. It holds a compressed, statistical sense of which businesses are associated with which categories, locations and attributes — built from directories, reviews, press, comparison articles and company sites. This layer favours businesses described consistently and often.
Retrieval. For most commercial questions, modern assistants also search the live web. They run one or more queries, fetch the top results, and read them. This layer behaves much like classic search: pages that rank for the assistant's internal queries get read. Pages that don't, don't.
Synthesis. The model composes an answer from what it retrieved and what it knows, deciding which sources to trust and which specifics to include. Pages that state things clearly and specifically are easier to synthesise from and more likely to be cited.
The practical consequence: you have no direct control over training, partial control over retrieval through SEO, and substantial control over synthesis through content structure and entity clarity.
Retrieval: being in the candidate set
Retrieval is where most businesses are lost before the answer is written. When an assistant searches, it typically runs informational and comparison-style queries — “best technical SEO agencies”, “technical SEO services UK”, “technical SEO agency reviews” — and reads a limited number of results, often fewer than ten per query.
That means classic ranking still decides your fate. If you're not on the first page for the queries the assistant runs, you're not retrieved and you can't be cited. The queries skew toward lists, comparisons and reviews, which is why third-party roundups and review platforms appear so often in AI citations: they rank for exactly those queries.
Implications: keep doing SEO; build pages that rank for comparison and “best for” queries; and earn presence on the roundups and review sites that already rank, because being in their lists gets you retrieved even when your own page isn't.
Trust: entity clarity and corroboration
Once retrieved, a source is judged. Models weight several things, most of which reduce to “is this a real, well-regarded entity, described consistently?”
- Entity consistency. If your site says one thing, your Google Business Profile another and a directory a third, the model hedges or omits you. Consistency across the web is a trust signal in itself.
- Corroboration. Claims that appear in more than one place — your site plus a review platform plus a press mention — are treated as more reliable than claims that appear only on your site.
- Authorship. Content from named, credentialed people is weighted more than anonymous marketing copy.
- Freshness. Dated, recently updated pages are preferred, particularly by retrieval-first engines.
- Reviews. Volume, recency and sentiment on Google and category platforms feed directly into recommendation answers.
Synthesis: being quotable
Finally the model writes. It prefers sources from which it can lift a clean, specific statement. Compare two pages for the same service:
Page A: “We deliver cutting-edge, results-driven technical SEO solutions tailored to your unique business needs.”
Page B: “Technical SEO covers crawlability, indexing, rendering, Core Web Vitals and structured data. Our audits take ten working days and produce a prioritised fix list; we implement on WordPress and Shopify directly.”
Page B is cited. It states what the service is, what it includes, how long it takes and what you get. Page A gives the model nothing to say. Specificity, structure and plain language are the core of GEO.
What each assistant rewards
| Assistant | Primary mechanism | What moves the needle |
|---|---|---|
| ChatGPT | Training + retrieval (search mode) | Broad, consistent presence across sources over time; ranking for comparison queries; quotable pages. ChatGPT SEO |
| Perplexity | Retrieval-first, always cites | Ranking for its internal queries; fresh, specific, well-structured pages; fast feedback loop. Perplexity SEO |
| Gemini | Google index + Knowledge Graph | Entity clarity, Business Profile, Google authority and reviews. Gemini SEO |
| Google AI Overviews | Google ranking + Gemini synthesis | Page-one ranking plus answer-first structure and schema. AI Overviews |
The differences matter at the margin. The core is shared: a consistent entity, specific answerable content, and presence on trusted third-party sources. Do those three well and you improve across every assistant at once.
What to do about it
- List your buyer questions and test them in each assistant. Record who is named and what is cited. This is your baseline.
- Fix your entity. Same name, description, category and location everywhere. Organization schema with sameAs links.
- Rewrite for specificity. Lead with the answer, state concrete facts, add structure and dates.
- Earn presence on the cited sources. Review platforms, directories, comparison articles in your category.
- Measure monthly. Mention rate, position, share of voice. Adjust.