How to appear in Google AI Overviews: the steps you control, and what only Google decides

The short version: nobody can make Google cite you in an AI Overview. Google decides whether an overview appears for a question at all, and whose pages it draws on. What you control is narrower: whether the machines assembling those answers can read your site, whether the facts on it are stated in a form they can use, and whether the rest of the web describes you specifically enough to be worth quoting. That is a finite list of work, with a check after each item — and it is the same work whether or not you end up cited.

An AI Overview is assembled, at the moment a question is asked, from the pages Google can read and trusts for that question. There is no placement to buy and no switch to flip. So this is not a list of tricks: it is the order we would work in on our own site, with an honest label on every step — prerequisite (nothing else works without it), lever (where the real gap usually sits, still no guarantee), or bonus (worth doing as hygiene; no evidence it moves Google). Our plain explanation of GEO covers how these answers get built, in more detail.

Step 1 — Measure first, before changing anything (prerequisite)

Why. "We're not in AI Overviews" is three different problems wearing one sentence. The machines may not be able to read you, the facts they read may be wrong or missing, or the answers may be built from sources that never mention you. Each needs different work — and on some questions no overview appears at all, so there is nothing to be in yet. Ranking well elsewhere proves nothing: in one local market we measured, the clinic was the most-named business in its market and was still cited in only 4 of its 12 AI Overviews — and 0 of its 25 AI Mode answers (the full report is public).

How.

1. Write down 10–20 questions a real buyer asks before they know your name — "best family dentist in Leeds", not "Acme Dental".
2. Put a few of them to Google the way a buyer would, in a clean session, and write the results down with the date. Note whether an overview appeared at all, and whose pages it quotes.
3. Open Search Console → Performance, where AI Overviews impressions are reported. It is the only view that comes from Google itself, and the one that shows which of your own pages its AI features actually surfaced.
4. Run the free check: it reads your site the way our measurement would, and shows you all 30 buyer questions in full. The method behind it is published, including what we can't measure.

Check: a dated baseline, and a first diagnosis — unreadable, misdescribed, or missing from the sources.

Step 2 — Fix the facts on your own pages (prerequisite)

Why. The answer is assembled from what can be parsed, and models reach for specifics. "Award-winning solutions for the modern enterprise" is invisible where "sports physiotherapy in central Bristol, evenings and Saturdays" is not. This is also the part you fully control — and the part most often missing. Price is the classic gap: the number is on the page, but nothing states in machine-readable form that it is the price of that thing, so the cost question gets answered by whoever did publish a parseable figure.

How.

• One page per service, saying what it is, where you do it, and your actual hours. Name things the way buyers say them, not the way your industry does.
• Put prices on the page, then mark them up — an Offer or price range in your structured data — so the figure is attached to the thing it prices.
• State the same facts the same way everywhere on the site. Contradictions read as a business that cannot be confirmed.

Check: our llms.txt checker flags a home page with no machine-readable offer, and the free check covers the rest of the technical layer.

Step 3 — Confirm the crawlers can read what you just fixed (prerequisite — the strictest one)

Why. Google's own documentation on its AI features says there are "no additional requirements … nor other special optimizations necessary" beyond being indexable and eligible for a snippet. That cuts both ways: there is no secret extra step, and there are several ordinary ways to be absent entirely. A stray noindex or nosnippet, or a site whose content only appears after JavaScript runs, can remove you from AI answers completely. This is the most common finding we see that is both severe and trivially fixable.

How.

• Check that your robots.txt does not block Googlebot, or the AI crawlers you care about. The llms.txt checker reports which crawlers your robots.txt blocks — free, no account.
• Open your page source and search for noindex and nosnippet. Staging directives left in production are common, and invisible on the rendered page.
• Turn JavaScript off and reload. If your services, prices and hours vanish, an assistant reading the HTML never saw them.
• If you have no /llms.txt, writing one is a bonus, not a prerequisite. These two small files (/llms.txt, /agents.md) tell a machine who you are — and on some store platforms they are generated for you, in which case they may introduce the platform and its checkout rather than your store. The checker reads both in seconds. Treat them as hygiene: Google has said its AI features do not read llms.txt.

Check: re-run the checker; the failure list should come back empty.

Step 4 — Get described where the answers are assembled (lever)

Why. This is where most missing citations live. In the market measured above, the sources the overviews leaned on were Google's own map listings first, then competing clinics' own pages, then the review and directory sites the answers depend on. None of them was the measured business's own site. To be cited, the answer's other sources have to know you too — your own claims about yourself carry the least weight of anything a model reads (how those answers get built).

How.

• Local businesses: claim and complete your Google Business Profile — precise category, named services, current hours, real photos. An empty field is a reason to skip you. Keep name, address and phone identical everywhere they appear; contradictions read as a business that cannot be confirmed.
• Ask satisfied customers to name the service in reviews. "Great place" carries nothing; "came in for Botox, loved the result" is evidence an assistant can quote — the cheapest visibility work there is.
• Find the directories and list pages that already rank for your category and get on them. A listing you submit yourself is usually free and takes ten minutes. For editorial "best of" lists, ask whoever maintains them what it takes to be considered.
• Do not assume the map pack covers this. The two systems read different sources and can disagree — we wrote up that measurement.

Check: our sample report prints the sources each answer was built from, question by question. That list is your work queue.

Step 5 — Write pages worth quoting (lever)

Why. An overview is a summary of sentences somebody wrote. A page that states its facts plainly and early can be drawn on; a page that makes a reader assemble the answer from three sections and a PDF cannot.

How.

• Put the answer in the first sentence, not the fourth paragraph.
• Use the buyer's question as the heading, then answer it completely underneath.
• Be concrete to the point of boring: amounts, ranges, areas served, response times, what you do not offer.
• Skip the volume play. Dozens of thin posts and mass directory submissions are the thing most likely to get you filtered, not the thing that gets you quoted.

Check: the lift test. Can you copy one sentence from the page and have it stand alone as a correct, complete answer to the question? If not, rewrite that sentence.

Step 6 — Re-measure and compare (prerequisite for knowing anything worked)

Why. Ask the same assistant the same question twice and the list can change; model updates reshuffle answers; whatever you measured is true for that day only. One run is an anecdote, not a measurement — and a before-and-after pair is the only thing that can tell you a fix did something.

How. Re-ask your 10–20 questions on a schedule — monthly is enough — the same way, in a clean session, and diff the answers against your Step 1 notes. Watch AI Overviews impressions in Search Console over the same window. If you would rather have software watch it, we compared the tracking tools and their published prices — but a dashboard watching a wrong answer only means you find out about the error every day instead of once.

Check: which fixes moved an answer, which did not, and which questions never changed because nothing was wrong to begin with.

Where the honest line is

Two things are true at once, and both need saying.

The prerequisites are real. If a page cannot be read, or its facts cannot be parsed, no amount of content strategy helps. That work is worth doing regardless of AI Overviews, and it is cheap.

And none of it is a guarantee. Google decides whether an overview appears and whose pages it cites. There is no schema tag that makes an assistant like you, there is no "GEO keyword", and no vendor can promise you placement. What you can give yourself is the honest position — crawlable, parseable, factually correct, worth quoting — and then the discipline to re-measure and see whether any of it moved.

One more honest note: this is not urgent for every business. If nobody would ever ask an assistant for a business like yours without already knowing your name, the whole field is currently irrelevant to you, and the money is better spent elsewhere. Everyone else: start where the diagnosis is cheapest. The free check reads your site, shows you every question we would ask, and takes about a minute. The full measurement runs all 30 questions across five AI surfaces — ChatGPT and Gemini with search on, Claude from memory, Google AI Overviews and AI Mode — and grades 176 answers from 302 API calls. $4.99, once.