Generative search, meaning Google's AI Overviews and AI Mode, ChatGPT search and Perplexity, composes an answer and cites a few sources instead of listing links. Being a source depends on three things: being indexed, having a passage that answers the sub-question directly, and having verifiable authority. Classic SEO is the prerequisite, not the competitor. Here are the mechanics, ten concrete actions and a measurement method.

Where generative search stands in 2026

  • Google AI Overviews launched in the US in May 2024 and reached more than a hundred countries the same year. For many question-shaped queries, a generated summary with source links sits above the results.
  • AI Mode, from 2025, is a separate tab where the whole result is a generated conversation with follow-ups.
  • ChatGPT search opened to everyone in late 2024 and relies largely on Bing's index plus OpenAI's own crawler for web retrieval.
  • Perplexity has been an answer engine with citations from the start, with its own crawler.

Coverage varies by language and market, but the direction is clear, and a site prepared now is ahead when the wave reaches its market. Since 2025 every SEO engagement we run includes a monthly 20-prompt panel that tracks exactly this.

How these systems choose a source

The internals are proprietary, but Google's documentation and observable behaviour agree on the shape:

  1. The query is fanned out into sub-questions. "Best framework for a multilingual e-commerce site" becomes separate retrievals about performance, payments, RTL support and so on.
  2. Pages are retrieved from the ordinary index. A page that is not indexed is never a source. Every rule in why a site is missing from Google applies unchanged.
  3. Passages are extracted, not whole pages. The model wants the paragraph that answers the sub-question directly and completely.
  4. The answer is composed and the passages used are cited. Sources that corroborate each other are cited more.

The practical consequence: the unit of competition is no longer the page. It is the paragraph. Each paragraph should be able to answer one question on its own.

What does not change from classic SEO

  • Crawling and indexing. If Googlebot cannot read the page, AI Overviews cannot either. Google states explicitly that no special markup or work is needed for AI features; the ordinary content guidelines apply.
  • Speed and mobile experience. A slow, cluttered page loses in retrieval too.
  • Domain authority. Links and mentions from other sites remain the primary authority signal.

So GEO is a layer on SEO, not a separate discipline. A vendor selling "AI SEO" as a package apart from technical SEO is charging twice for one job.

Ten concrete actions

  1. An answer block under every heading. The first paragraph after the H1, and after every question-shaped H2, answers completely in 40 to 60 words. This article does exactly that.
  2. Question headings in the user's words. "How much does a RAG chatbot cost" beats "Cost analysis of conversational solutions".
  3. One page, one subject. A page that half-covers five topics is the best passage for none of them.
  4. Numbers with dates. "Corporate site from 35 million toman in 2026" is quotable; "competitive pricing" is not.
  5. A real author and an about page. Models and crawlers build an entity from name, organisation, address and external profiles. Organization and Person schema with sameAs help.
  6. A real updated date. A page refreshed quarterly with new figures wins time-sensitive questions. A fake date does the opposite.
  7. Comparisons and decision lists. "Which is better" questions are the largest share of generative queries. Honest comparison against criteria, not advertising.
  8. Mentions on other sites. Directories, reviews, interviews, guest articles. When a model sees your name in several independent sources, it treats you as a more credible answer.
  9. Delete the filler. A three-paragraph "in today's fast-moving world" opener buries the answer and lowers the chance of the passage being selected.
  10. Correct structured data. FAQPage, Article with author and datePublished, Product with a real price. Schema does not rank pages by itself; it makes extraction more precise.

llms.txt, robots.txt and AI crawlers: the decision

  • llms.txt is a proposed file that summarises a site for language models. Google has said it does not use it, and evidence of use by other engines is thin. Publishing one is harmless; counting on it is a mistake.
  • AI crawlers have distinct names: Google-Extended (training only, not search), GPTBot and OAI-SearchBot, ClaudeBot, PerplexityBot. Blocking Google-Extended has no effect on AI Overviews, which draw on the ordinary Google index.
  • Our recommendation for a business site: leave search and citation crawlers open, because being cited is the goal. If training on your content is the concern, block only the training crawlers. A publisher whose revenue is the traffic itself faces a different equation.

How to measure it

  • Search Console. Since 2025 Google counts AI Mode and AI Overview clicks and impressions in the performance report, but does not separate them. The practical tell: flat impressions with falling clicks on question-shaped queries.
  • Referrals in analytics. Define chatgpt.com, perplexity.ai and copilot.microsoft.com as their own channel. Small still, but the conversion rate is usually high.
  • A monthly prompt panel. The 20 questions your real customers ask, put to four systems every month, recording which sources were cited. This is the only direct method, and we run it for every SEO client.
  • Brand search. When you are cited, searches for your company name rise. Record it monthly.

The opportunity for a niche or non-English site

High-quality sources are scarce for most technical and commercial subjects outside English. A model looking for a good passage on "payroll software cost in Iran" has few candidates. A page that answers with figures, a date and a real author becomes a source with a fraction of the competition English faces. That logic drives our own blog schedule: one real question, one complete answer, one real project per article. Our commercial pages follow the same structure: question-shaped sections, prices with a year, FAQ with schema.

Do AI Overviews reduce traffic?

For simple informational queries, yes; the user gets the answer in place. For commercial and complex queries, the click from being cited is higher quality than before, because the visitor arrives with initial trust. The response is to become the source, not to resist.

Should we produce more content?

No, more precise content. Ten pages that each answer one question completely are cited more than a hundred general pages. Mass-produced AI content does the opposite and collides with Google's scaled-content-abuse policy.

How long does GEO take?

If the site is indexed and healthy, restructured answers show up in the prompt panel within weeks. External authority, mentions and links, takes months, as it always has.

Summary

Generative search moved the unit of competition from the page to the paragraph. With crawlability, speed and authority in place, the rest is answer structure: question headings, answer blocks, dated figures, a real author. We measure it with a monthly prompt panel, and if you want to know whether your site appears in AI answers today, we will run the same 20 prompts for you.

Sources