Google AI Overviews optimization and AI Mode

For a lot of queries, Google now answers before it lists. The generated block at the top takes the click, and the pages it cites are not always the pages ranking underneath it. This page covers the part of our GEO services aimed at AI Overviews and AI Mode, which are the two places that answer appears.

How do these answers get built?

By splitting your question into several.

Instead of matching one query against one index, the system fans the query out into related sub-questions, runs each of them, and assembles an answer from what comes back. Ask about choosing a supplier and it may separately look up typical prices, common contract terms, how long the work takes and what usually goes wrong, then stitch the pieces together.

That is why the citation list often includes pages that do not rank in the top ten for the original question. They rank for one of the smaller questions instead. It also explains why a thorough page covering price, timeline, deliverables and failure modes tends to be cited more than a page that answers only the headline question well.

Does ranking in the top ten get you in?

It helps, and it is not the whole story.

The overlap between the pages cited in generated answers and the organic top ten falls well short of complete, so a strong ranking is a good start and not a ticket. Because the system asks several smaller questions and fetches for each one, a page can be cited for a sub-question it answers well while ranking nowhere for the headline query.

So the plan has two halves. Cover the sub-questions properly on your own pages, so there is something to fetch for each of them. Then be visible enough elsewhere that Google's index has more than one route to you. Links still matter and we still earn them, and so does being written about by people who are not you.

Which crawler controls this?

Googlebot, and only Googlebot.

This is the most commonly misunderstood rule in the whole field, so it is worth being blunt. Google-Extended does not control AI Overviews or AI Mode. Blocking it removes you from training for the Gemini apps and Vertex, and changes nothing about whether you are cited in a generated Google answer. The only way to leave AI Overviews is to leave Google search.

The other engines read other indexes and want other crawlers, which is why opening Googlebot does nothing for ChatGPT, Perplexity or Claude. The full six-engine comparison sits on the GEO services page.

What do we change on your pages?

We build for the sub-questions, not the headline.

For each target topic we map the smaller questions the system is likely to fan out to, then make sure one page answers all of them, each under its own heading, each in a passage short enough to lift. Price ranges, timelines, what is included, what is excluded, what usually goes wrong. Numbers get named sources and dates.

We also check what Googlebot actually receives. Text that only appears after a script runs, redirect chains, slow templates and markup that contradicts the visible page all cost you here in the same way they cost you in normal rankings. That is ordinary technical SEO work, and it is the floor everything else sits on.

How do you see any of this in your own data?

Through first-party reports, which finally exist.

Aug 31, 2026

That report shows impressions and clicks for generated answers, rather than leaving you to infer a change from a dip in the totals. Bing's tools do the same for per-page Copilot citations. Those two are the only numbers that come from the engines themselves, so we anchor the reporting on them and use our own prompt tracking to explain the movement.

What does the work look like?

Which pages do we work on first?

The ones already losing clicks to an answer.

The first pass compares your Search Console data against which queries now return a generated block. Pages holding their position while their clicks fall are the clearest opportunity, because the demand is proven and the answer above them is taking it. Those get rewritten to cover the sub-questions properly, so you have a chance of being inside the block rather than under it.

Second come the queries where a generated answer appears and you are absent from both the answer and the top ten. Those are slower, and we scope them honestly rather than promising a category question in a quarter. Last come pages that still convert well with no generated answer above them. They are working, and the fastest way to break something is to rewrite it for a problem it does not have.

What does not help here?

Publishing an llms.txt file. No answer engine uses it to decide citations, Google said in June 2026 that it has no effect, and a study of 137,000 domains found 97 percent of them were never requested.

Adding FAQ markup to chase a rich result. Google retired FAQPage rich results in May 2026. We keep accurate markup where it mirrors visible text and we keep the visible questions, because the questions always did the work.

Snippet limits. Telling Google how little of your page it may show also tells it how little it may quote, which is the opposite of the goal on a page you want cited.

The neighbouring pages here are Claude optimization and the AI visibility audit. To see how we run this for a team in one of our regional markets, our market page walks through the same process.

How do we start?

Book a free 30-minute consultation. We check crawler access, whether your key text exists in the raw HTML, and what your markup says, then name the first three fixes.

Plans are flat monthly retainers with starting prices on the pricing page. They run month to month, and you can stop at the end of any month.