AI Search Optimization: How to Get Your Website Cited by ChatGPT, Perplexity and Google AI Overviews

The short version: AI assistants answer questions by retrieving a handful of passages from the open web and quoting them. To be one of those passages, your page needs to answer a specific question in plain language, near the top, in a self-contained block that still makes sense when it is lifted out of its surroundings. Everything else in this article is detail on that one idea.

For fifteen years the job was to rank. You wrote a page, Google put it in an ordered list, and if you were in the top three you got the click. That still works — but it is no longer the whole picture. A growing share of searches now end in a generated answer, with three or four sources cited underneath it, and no visit to anyone's website at all unless the reader chooses to dig deeper.

Reported figures vary by source and by how you count, but the direction is consistent: Google's AI Overviews are estimated to appear on roughly 30–40% of queries, ChatGPT is reported to serve on the order of 200 million weekly users, and Perplexity around 100 million queries a month. Treat the exact numbers as moving targets. The structural change underneath them is the part worth planning around.

This shift has picked up a name — generative engine optimization, or GEO — along with the usual crop of people selling it as a brand-new discipline. It mostly is not. Most of what makes a page citable is what made it good before: it is fast, crawlable, well-structured, and it actually answers something. What follows is the part that genuinely differs, and what to change on your site.

How AI Engines Actually Choose Sources

It helps to know roughly what happens between someone's question and a cited answer. The details differ by product, but the shape is broadly the same:

• The question gets rewritten. One conversational question — "what should I budget for a new website for my plumbing business?" — is expanded into several conventional search queries behind the scenes.
• Candidate pages are retrieved using a conventional search index. This is the important part: if you cannot be found by search, you cannot be cited. Classic SEO is the entry ticket, not an alternative strategy.
• Pages are split into passages. The model does not read your page as a page. It reads chunks — a heading with the paragraphs beneath it, a list, a table row.
• The most relevant passages are assembled into a context window, and the model writes an answer grounded in them, attaching citations to the sources it leaned on.

The practical consequence of step three is the one most people miss. Your competition is not other websites. It is other paragraphs. A brilliant 3,000-word guide whose key point only makes sense after 2,000 words of preamble will lose to a competitor's clean 60-word definition, because the definition survives being cut out of its page and the guide does not.

1. Answer The Question In The First 40 Words

Under every H2, lead with a direct, complete answer, then elaborate. Not "in this section we'll explore the factors affecting cost" — instead, "A five-page business website typically costs between X and Y, driven mainly by content volume and custom interaction work."

Write each answer so it survives extraction. That means avoiding pronouns that depend on the previous paragraph ("as we mentioned above", "this approach"), and naming the subject explicitly even when it feels slightly repetitive to a human reading top to bottom. A passage that begins "It usually takes four to six weeks" is useless out of context. "A five-page Webflow build usually takes four to six weeks" can be quoted anywhere.

2. Use Question-Shaped Headings

People type full questions into AI assistants far more often than they typed them into Google. Headings phrased as questions — "How long does a Webflow build take?" rather than "Timelines" — match those queries more directly, and they force the paragraph underneath to be an answer rather than a meander.

This is also where a genuine FAQ section earns its place. Not the padded kind written for a keyword, but six or eight real questions you get asked on sales calls, answered honestly and specifically. Those blocks are close to ideal retrieval units: short, self-contained, and unambiguously matched to a question.

3. Give Them Something Specific To Quote

Generative engines cite passages that carry information the model cannot produce on its own. Vague, hedged marketing prose is exactly what it can already produce, so there is no reason to attribute it to you. What earns a citation:

• Numbers with context — prices, timelines, measured results, percentages, sample sizes.
• Dates and versions — "as of August 2026" tells a retrieval system your content is current, and gives the model something to anchor a freshness claim to.
• Named entities — specific tools, standards, integrations, and platforms rather than "modern technologies".
• Original data — even small first-party numbers, like the average load time across the last twenty sites you built, are things no one else can supply.
• A clear position — an honest "we do not recommend X for Y, because Z" is far more quotable than balanced-on-all-sides copy.

The uncomfortable corollary: if your page contains nothing a language model could not have written from general knowledge, it will not be cited, no matter how well optimised the markup is.

4. Make The Structure Machine-Legible

Clean semantic HTML matters more than it did, because chunking follows document structure. Concretely:

• One <h1>, then a logical <h2>/<h3> hierarchy that does not skip levels.
• Real <table> markup for tabular data, not columns faked with divs. Tables are unusually citable — comparison rows get quoted verbatim.
• Real <ul> and <ol> lists rather than paragraphs of manually typed bullets.
• Descriptive alt text, since images themselves contribute nothing to a text retrieval pipeline.

Add structured data on top: Article or BlogPosting with a real author and dates, FAQPage for question blocks, Organization and LocalBusiness on the pages that describe your company, Product or Service where relevant. Schema is not a magic ranking input, but it removes ambiguity about what your page is and who stands behind it — and that is a real input to how confidently a system will attribute something to you.

5. Render Your Content In HTML, Not JavaScript

Googlebot executes JavaScript. Many AI crawlers do not, or do so inconsistently. If your key content only appears after a client-side framework hydrates, there is a meaningful chance the crawler sees an empty shell.

The test takes ten seconds — fetch the page without running scripts and see what is actually there:

curl -s https://yoursite.com/your-page | grep -c "a distinctive sentence from your page"

If that returns 0, your content is invisible to anything that does not render JavaScript. This is one of the quieter advantages of static HTML and of Webflow's published output — the content is simply in the document, with nothing to execute first.

6. Speed And Crawlability Still Apply

Core Web Vitals remain a confirmed Google ranking signal, and since AI Overviews are built on Google's own index, the same fundamentals carry through. As of 2026 the "good" thresholds, measured at the 75th percentile of real visitor data over a rolling 28-day window, are:

INP is the one most sites fail — reporting through 2026 has consistently put the failure rate above 40%, largely because it is driven by heavy third-party scripts, chat widgets, and tag managers rather than by anything in the design. It is worth auditing what you have loaded on every page before optimising anything else.

Beyond speed, check the boring plumbing: an accurate XML sitemap, a robots.txt that does not accidentally block what you want indexed, canonical tags on every page, and no orphan pages sitting outside your internal link graph.

7. Decide Deliberately Whether To Allow AI Crawlers

Different bots do different jobs, and conflating them is a common and expensive mistake. The main ones to know:

Two things worth being precise about. First, Google-Extended does not control AI Overviews. Overviews are served from the standard Google index via Googlebot, so opting out of Google-Extended will not remove you from them — the only way out is to leave Google's index entirely, which is almost never what a business wants. Second, blocking the search-and-citation bots is a different decision from blocking the training bots. If you are a business that wants to be found, being absent from the assistants your buyers are asking is a strange thing to pay for.

Publishers with licensing leverage may reasonably conclude otherwise. Most service businesses should not.

8. About llms.txt

You will see llms.txt recommended a lot right now. Being straight about it: it is a proposed convention — a markdown file at your root that summarises your site and links its important pages — and no major AI engine has publicly confirmed it as a retrieval or ranking input.

It takes twenty minutes to write and does no harm; we publish one at webnovex.com/llms.txt. Just file it under cheap housekeeping rather than a growth lever, and be suspicious of anyone selling it as the centrepiece of an AI search strategy. If you had to choose between writing an llms.txt and rewriting your six most important pages to lead with direct answers, the rewrite wins by a distance.

How To Tell Whether Any Of This Is Working

Attribution is genuinely harder here than in classic SEO, because a citation that never gets clicked leaves no trace in your analytics. Three things to watch:

• Server logs. Filter for the user agents in the table above. This tells you who is crawling you and how often — the leading indicator, before any citations appear.
• Referral traffic. Watch for chatgpt.com, perplexity.ai, copilot.microsoft.com and gemini.google.com in your referrers. Volumes are small compared with organic search, but this traffic tends to convert well — the reader has already been told you are a credible answer.
• Manual prompt testing. Write down fifteen questions a real buyer would ask, run them across ChatGPT, Perplexity and Google every month, and record who gets cited. Unglamorous, and still the most honest signal available. Commercial AI-visibility trackers exist, but the category is young and coverage is uneven.

Also expect your Search Console click-through rate to fall on informational queries even while impressions hold steady. That is the shift working as designed, not a penalty. Judge informational pages on citations and assisted conversions; keep judging your commercial pages on clicks.

What This Means For Your Site

Nothing here requires abandoning what you have. In order of return on effort:

• Take your six highest-intent pages and rewrite the opening 40 words under each heading to answer the question directly.
• Convert vague headings into the questions your customers actually ask.
• Add the specifics you have been hedging on — real prices, real timelines, real results.
• Add Article, FAQPage and Organization schema.
• Confirm your content exists in the raw HTML.
• Fix whatever is failing INP, which is usually a third-party script you forgot about.
• Check your robots.txt is not quietly excluding you from the assistants your buyers use.

That list is a couple of weeks of focused work for most business sites, and every item on it makes the page better for human readers too. That is the useful test when someone tries to sell you an AI search package: if a tactic only helps machines and makes the page worse to read, it is not a strategy — it is this decade's keyword stuffing, and it will age the same way.

If you want a hand applying this to your own site, that is a large part of what we do — see Webflow website development and website redesign, or read our companion guide on how to rank a website on Google for the conventional SEO groundwork this all sits on top of.