AI search optimization: how to get cited by ChatGPT and Google AI

For one month in 2025, Pew Research watched what 900 American adults did with Google. Across 68,879 searches, 18% returned an AI summary. When one appeared, people clicked a search result on 8% of visits. When it did not, they clicked on 15%. Ahrefs found the same effect from the publisher's side, putting the click loss to the top-ranking page at 34.5% across 300,000 keywords.

AI search optimization is the work of staying visible when the summary is the answer. It covers Google's AI Overviews and AI Mode, plus ChatGPT, Perplexity, Copilot and Gemini, which run their own crawlers and their own indexes.

Most of the advice written about it is older than the docs. Google's May 2026 guide to optimizing for AI features devotes a section to the things you do not need to do. It names answer engine optimization and generative engine optimization, then knocks down llms.txt files, content chunking, AI-specific rewriting and bought mentions.

Set your expectation before you spend. A citation puts your name in front of somebody who is deciding, and it rarely sends a click. Judge it on inquiries, not on sessions.

Less than the category name suggests. Google says its AI features run on the core Search ranking and quality systems, and that there are no additional requirements to appear in them. A page Google cannot index is a page Google cannot sum up, so ordinary technical SEO still comes first.

Six differences are worth knowing, and most of them sit outside Google.

Which index each assistant queries decides who is even in the running. We covered that plumbing in a separate article on how AI assistants choose sources, and Google's two surfaces get their own guide. This one stays on execution.

Start at the firewall, because this part is binary and the platforms document it themselves. OpenAI runs three agents, and owners often block the wrong one. GPTBot collects training data, OAI-SearchBot surfaces sites in ChatGPT search, and ChatGPT-User fetches a page on request. OpenAI says sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers.

Blocking GPTBot to keep your writing out of a training run is a fair call. Block every OpenAI agent to get there and you also drop out of ChatGPT search answers, with no error message. Perplexity draws the same line: PerplexityBot indexes pages for search and is not used for model training.

Cloudflare compared May 2024 with May 2025 and found Googlebot's share of crawler requests up from 30% to 50%, GPTBot up from 2.2% to 7.7%, and ClaudeBot down from 11.7% to 5.4%.

Two settings can remove you without anyone touching robots.txt. Google says that nosnippet, data-nosnippet, max-snippet and noindex set your display preferences, and that stricter ones limit how AI features can use your content. A max-snippet of zero, left on a template after an old scraping scare, blocks every page built on it. Google also asks that a site be included in Search generative AI features in Search Console.

The strongest evidence here is a paper rather than a blog post. Researchers built GEO-bench, 10,000 test queries across 25 subject areas, tried nine ways of editing a page against an unedited control, and published the results at KDD 2024. The winning edits were adding direct quotes, adding statistics and citing sources, which the paper puts at a 30 to 40% relative gain.

Two results deserve a place on the wall. Keyword stuffing scored 17.7 against a do-nothing baseline of 19.3, so it left pages less visible than leaving them alone. Edits to fluency and reading level, which change no facts, lifted visibility by 15 to 30%. Tested on Perplexity, a live engine, the same methods lifted visibility by up to 37%.

That reads like an editing brief, so treat it as one.

Google puts the same idea above every other step in its guide: pages built on what you have seen and done. Local firms should keep the Google Business Profile current too, since Google feeds it into AI answers about local businesses.

No evidence says it does. Jeremy Howard proposed llms.txt in September 2024 as a plain-text file telling AI systems where a site's important content lives. The proposal claims no adoption by any AI company.

The field data is worse than that modesty suggests. Ahrefs read server logs across 137,210 domains and published the numbers in June 2026. Of those domains, 28% publish an llms.txt file, and 97% of the files got zero requests in May 2026. AI retrieval bots, the only kind that could cite you, made up 1.1% of the AI bot requests to files that got any traffic. The study also found zero requests from AI bots for llms.txt files that do not exist, so the systems never check for one. Google says Search does not use llms.txt or any other special machine-readable file.

Three layers exist, and the first two are free. Google Search Console launched Generative AI performance reports on June 3, 2026, showing impressions in AI Overviews, AI Mode and Discover by page, country, device and date. Read the caveats with the numbers: impressions and not clicks, more metrics promised over time, and a rollout that reached a subset of sites first.

The second layer is GA4, where you segment referrals from chatgpt.com, perplexity.ai, gemini.google.com and copilot.microsoft.com. Several assistants strip the referrer, so treat that figure as a floor. Our GA4 reference build covers the setup. The third free layer is your own server logs, the only place that confirms a crawler reached a page.

Before any of that, take a baseline you did not pay for. Our free SEO report scores any URL on ten weighted on-page signals as an AI Visibility Score, beside Core Web Vitals and the usual checks. It reads your page. It cannot see whether ChatGPT cited you last Tuesday, and no on-page tool can.

Paid prompt trackers answer a different question: which of your buying questions surface you, and who gets named instead. Work the free layer first, then decide whether a monthly tracker earns its line in the budget. Checked every month rather than once, that becomes a standing program, which is the shape our AI visibility service takes.

The list of things not to do is short, well documented and expensive to ignore.

• Keyword stuffing for AI, which the KDD 2024 benchmark measured at 17.7 against a 19.3 baseline for changing nothing.
• Chopping pages into fragments, since Google says there is no requirement to break content into tiny pieces.
• Rewriting in a stiff AI-friendly voice, since Google says you do not need to write a specific way for AI search.
• Buying mentions, which Google calls inauthentic and less helpful than it might seem.
• Making a page for every question variant, which Google says breaks its scaled content abuse spam policy.
• Treating schema markup as a citation lever, since Google says there is no special structured data you need to add for AI features. Keep your schema accurate because it earns rich results in ordinary search.
• Buying a monitoring platform at several hundred dollars a month before anyone has read your robots.txt file. A tool reports that you are missing, not that a firewall rule is why.

Run it in this order, because the cheap work gates the expensive work.

1. Audit crawler access and snippet settings. One hour, and it decides whether anything else counts.
2. Turn on the Generative AI performance report in Search Console and write down the number you start with.
3. Edit your ten best pages against the checklist above: sources named, numbers dated, answers first.
4. Baseline ten to twenty buying questions. Run them through ChatGPT and Perplexity monthly and record who gets named.
5. Fix the structure and reading level of those pages, worth 15 to 30% in the benchmark.
6. Publish what no competitor can copy: your own numbers, your own jobs, the case that went sideways.

Pew's 900 adults clicked a link inside an AI summary on 1% of the visits where one appeared. Work through that list and you should show up in more summaries while your traffic line barely moves. Ask each new caller where they heard about you, keep the answers for a quarter, and read them beside the impression count.