Skip to content
Aquinel
Free Consultation Call, No Obligation

How to Optimize for AI Search: Your Website and Content, Step by Step

How to optimize for AI search: let the crawlers in, match each page to a real question, answer it first, add something only you have, and measure beyond Google.

Lourdes Paul Agilan, founder of Aquinel

Founder, Aquinel · Published · 14 min read · Updated

How to optimize for AI search: make your website easy to crawl, easy to read and worth quoting. Let AI crawlers in through robots.txt, keep key content as real text, match each page to the exact question a buyer asks, answer it in the first sentence, add something only you have, then measure in Search Console and by prompting the AI tools yourself. Most of this is good SEO.

That last sentence is the part people do not want to hear, so let me be straight up front. If you want to know how to optimize for AI search, most of the tactics look a lot like SEO. Three things genuinely do not: how you measure it, how you research prompts, and the fact that Google is only one of the places you need to show up. If the definitions are still fuzzy, here is what generative engine optimization means.

What AI search does with your page

Three things happen before your content can be quoted.

The question gets split up. ChatGPT, Google AI Overviews and AI Mode all rewrite a question into several narrower searches before answering. Google calls this "query fan-out" (Google Search Central, updated 10 December 2025). Your page competes on those sub-questions, not just on what the person typed. We cover the details in how ChatGPT search works and how to appear in Google AI Overview.

The page gets cut into pieces. Microsoft's Bing team says assistants like Copilot "break content down, a process called parsing, into smaller, structured pieces" (Microsoft Advertising, 8 October 2025). The model quotes a piece, not a page.

Often, only the top gets read. RESONEO, a French SEO consultancy, found that in ChatGPT's instant mode the model usually sees a snippet of about 200 characters from the H1 and the text right after it. Pages ChatGPT actually opened were cited 74% of the time, against 7% for pages it only retrieved (Search Engine Land, 17 August 2026).

Step 1: Let the crawlers in

Nothing else matters if the bots cannot read your pages.

Open your robots.txt and look for the AI crawlers by name. OpenAI runs four, and two matter most. OAI-SearchBot surfaces your site in ChatGPT's search answers. GPTBot collects content for training. Blocking GPTBot keeps you out of training data but leaves ChatGPT search alone. Blocking OAI-SearchBot takes you out of ChatGPT search results. Check your CDN and firewall rules too.

Google says there are no additional requirements to appear in AI Overviews or AI Mode. Your page needs to be indexed, eligible to show with a snippet, and crawlable. The one thing Google stresses is that important content has to exist as text.

That gate is real. RESONEO found ChatGPT's fetcher does not run JavaScript, so text that only appears after scripts run is invisible to it, and pages over 4 MB are rejected outright. Microsoft advises against hiding "important answers in tabs or expandable menus," and notes HTML gives better structure signals than PDFs. Keep key details out of images, or at least in alt text.

One more box. Since 31 August 2026, Search Console has a Search generative AI control that keeps your content out of Google's AI features. Confirm nobody switched it on. Google says it is not a ranking factor for ordinary search.

Step 2: Read what Google actually published

In May 2026 Google published a guide on optimizing for generative AI features. It is worth twenty minutes, mostly for what it tells you to stop doing.

Google's position is that "optimizing for generative AI search is optimizing for the search experience, and thus still SEO." It says you do not need new machine-readable files, AI text files, markup or Markdown versions of your pages. You do not need to break content into tiny pieces or write in a special way for generative engines. Chasing inauthentic mentions does not help.

Ethan Lazuk, writing the day after, kept llms.txt on his own site, arguing there is no real downside. My read: it costs an hour and has no published evidence behind it, so treat it as a cheap experiment, not a step.

Step 3: Match each page to the question

Discovered Labs, an AI search agency, analysed about 2 million citations across 10,000 cited pages on ChatGPT, Claude, Google AI and Gemini over six months. Its strongest page-level signal was how closely the page matched the question asked. The effect (+0.37) was roughly three times the next page-level signal, and one standard deviation more alignment meant about 30% more citations (Discovered Labs, updated August 2026, vendor-funded).

Ahrefs found the same thing from another angle. Across 1.4 million ChatGPT prompts, cited page titles were closer in meaning to ChatGPT's rewritten sub-queries (0.656) than non-cited titles were to the original prompts (0.484). Pages with plain-language URL slugs had an 89.78% citation rate, against 81.11% for pages without (Ahrefs, 15 April 2026, vendor-funded).

What this means in practice:

  • Start from real questions, not keywords. Pull them from sales calls, support tickets and reviews, in the buyer's words.
  • Give each important question its own page or section. A page that answers "What does X cost for a 50-person team?" is a better match than a general pricing page.
  • Put the question's words in the title, the H1 and the slug.
  • Cover the pages AI now searches for on your own site. Promptwatch, which sells AI visibility tracking, measured searches scoped to named websites jumping from 0.37% to 16.8% of ChatGPT's sub-queries on 8 August 2026 (vendor-funded). When a model searches your site for "pricing" or "integrations", you need a page that answers it.

Step 4: Answer in the first sentence

Two studies using different methods point the same way.

Kevin Indig matched 18,012 verified ChatGPT citations back to their source text. 44.2% came from the first 30% of the page, 31.1% from the middle and 24.7% from the final third (Search Engine Land, 18 February 2026). Adam Gnuse audited 15 domains getting real ChatGPT referral traffic and found 72.4% of cited blog posts had an "answer capsule": a self-contained answer of about 20 to 25 words right after a title or question-shaped heading, almost always with no links in it (Search Engine Land, 19 November 2025).

So put the question as the heading, answer it fully in the next sentence, keep links out of that sentence, then explain. Breadcrumbs, bylines and share buttons between your H1 and that answer use up the few hundred characters a model may see first.

Then write the rest in pieces that work alone. Indig found definite statements ("X is", "X refers to") were cited more than hedged ones, and question headings made content about twice as likely to be cited. Cited passages were about 20.6% proper nouns, against 5% to 8% in typical text. Replace "it" and "this" with the actual name. Our guide on how to structure content for AI goes deeper.

Step 5: Improve the page, not the tricks

The original 2024 GEO paper, on a 10,000-query test set, found adding quotations lifted visibility by about 41%, statistics by about 31% and citing sources by about 27%. Keyword stuffing made things about 8% worse (Aggarwal et al., KDD 2024).

A newer paper pushes back. Liu and Xu, in April 2026, tested those same tactics on GPT-4o-mini and found every one scored below an untouched page. The unmodified page hit 13.34% visibility. Statistics dropped it to 12.21%, quotations to 11.08%, citing sources to 11.78%. Their own method, which reworks the whole document rather than bolting on phrases, reached 18.31%. Their explanation rings true: the additions "introduce redundancy and disrupt the naturalness signals that generative engines implicitly favor."

Nobody has reconciled the two papers. Stuffing a statistic into a thin page gives you a thin page with a statistic in it. So add something genuinely yours: numbers from your own customers or tests, a named method with its steps, a quote from someone who did the work, or a clear position with the reasoning shown. Gnuse found 52.2% of cited posts had original data or brand-owned insight.

One warning from Aquinel (then GeoCited). Seventeen days after we published original numbers, Google's AI Overview still used our figure but credited a LinkedIn post about the study and another site carrying our title (how to appear in Google AI Overview). Put your brand name in the same sentence as your key number.

Step 6: Choose the format from the question

The right format depends on the question. In our study of 40 buying-intent prompts across eight B2B software categories, run on ChatGPT and Google AI Overviews on 28 August 2026, the winning format changed with intent (which content formats win AI citations):

Question typeFormat that took the most citationsShare
Head-to-head ("X vs Y")Comparison pages56.7%
PricingPricing pages53.4%
Selection ("which X for my team")Product pages32.8%
Category ("best X")Listicles30.0%

Wix Studio's AI Search Lab, using Peec AI data (both vendors), found listicles made up 35.37% of citations in its SaaS group (Wix Studio, 23 March 2026). Our listicle share was 12.9% overall. The likely reason is the question mix: 40% of our prompts were comparison or alternatives questions, and Wix does not publish its mix. Either way, look at the questions your buyers ask, then build the format those questions reward.

Step 7: Update when something changes, not to change the date

The evidence is mixed. Ahrefs, across 16.975 million citations, found AI-cited pages were 25.7% fresher than organic results, with ChatGPT preferring newer pages most (Ahrefs, 28 July 2025, vendor-funded). Its later study put the median page ChatGPT cited at around 500 days old (Ahrefs, 15 April 2026). Freshness helps where the answer changes, like prices or features. Update when the facts change. Changing only the date adds nothing.

What the evidence doesn't support

Schema markup as a citation tactic. Ahrefs tracked 1,885 pages that added JSON-LD between August 2025 and March 2026 against 4,000 matched controls. Citations fell 4.6% in AI Overviews, a small but significant drop, and rose 2.4% in AI Mode and 2.2% in ChatGPT, both inside the noise (Ahrefs, 11 May 2026, vendor-funded). Otterly, also a vendor, found six of seven platforms could not fetch or correctly read schema. Microsoft still recommends structured data. Add it for ordinary search and keep it matching the visible text.

Word count targets. Ahrefs found 53.4% of pages cited in AI Overviews had fewer than 1,000 words, with almost no link between length and position (Ahrefs, 3 December 2025, vendor-funded). Write as long as the question needs.

Keyword stuffing. It was the only method that hurt results in the original GEO research.

Good content on a weak domain. Discovered Labs found how much AI systems seemed to trust the domain was about six times more influential than the strongest page-level feature. Mentions of your brand on other sites still matter. See how to build topical authority for AI.

Step 8: Turn measurement on, then learn to distrust it

Search Console now has a Generative AI performance report covering AI Overviews and AI Mode, rolled out worldwide on 31 August 2026. It is the only first-party number you will get. It reports impressions only: no clicks, no click-through rate, no query data. An impression means Google showed a link to your site inside an AI feature, not that anybody visited.

Read it carefully. From 13 August 2026, a logging error made impressions in that report fall off a cliff. John Mueller confirmed it was "just a logging issue and not representative of visibility changes in Search." Before you change anything based on a drop, check your server logs for AI crawler hits and compare against ordinary organic clicks for the same pages.

Step 9: Look outside Google

The Search Console report is Google only. It tells you nothing about ChatGPT, Perplexity, Claude or Copilot.

So build a small prompt set, twenty or thirty questions a real customer would ask, and run them by hand across the platforms you care about. Record who gets named and which pages get quoted. Do it monthly, on the same day, and judge results over several weeks, not from one answer. Watch your server logs too: hits from GPTBot, PerplexityBot and Google-Extended tell you who is reading you, and they move before the dashboards do.

  1. Check robots.txt, CDN rules and the Search Console opt-out.
  2. Keep key content as HTML text, visible without clicks or scripts.
  3. Put each page's exact question in its title, H1 and slug.
  4. Answer it first, in about 20 to 25 words, with no links.
  5. Add one thing only you have, with your brand name next to it.
  6. Pick the format the question rewards.
  7. Watch the Generative AI report for a month, then run your own prompt set.

If you would rather not squeeze that between other work, that is what our services cover.

Questions

What is the short version of how to optimize content for AI search?

Match the page to the exact question a buyer asks, answer it in the first sentence under a question heading, add original data, and keep the text in plain HTML. In a vendor study by Discovered Labs of about 2 million citations, matching the question had about three times the effect of the next page-level factor.

Is AI search optimization different from SEO?

The tactics mostly are not. Google says optimizing for generative AI search is still SEO, and the technical and content work looks familiar. What genuinely differs is measurement, prompt research, and covering platforms outside Google that publish no data about you.

Is there a special trick for how to optimize content for LLMs?

No. Google says you do not need to write in a special way for generative engines or create AI text files. A 2026 paper by Liu and Xu found that bolting quotations, statistics or citations onto a page scored below the untouched page. Writing a genuinely better page did better.

Will adding schema markup get me cited more by AI?

Probably not directly. Ahrefs tracked 1,885 pages that added JSON-LD against 4,000 controls and found citations barely moved, with a small drop in Google AI Overviews. Otterly found six of seven AI platforms could not read schema correctly when asked. Both are vendors. Add schema for ordinary search and keep it accurate.

How do I track whether AI search is sending me traffic?

Search Console's Generative AI report gives you impressions for AI Overviews and AI Mode, but no clicks. For actual visits, segment referral traffic from AI domains in your analytics. For ChatGPT and Perplexity visibility you have to run your own prompts, because no first-party report exists.

Sources

If you would like us to walk you through how this would work for your firm, the call is free.

Lourdes Paul Agilan, founder of Aquinel

About the author

Lourdes Paul Agilan

Founder of Aquinel. Works with MSPs and IT services firms from Chennai, India, on getting recommended by ChatGPT and found on Google.

LinkedIn

Keep reading

See how this would work for your firm.

Free Consultation Call, No Obligation