How to Get Your Website Found by ChatGPT, Gemini and Google AI

Assistants and classic search engines do not share one ranking formula, but they do share a dependency: they can only use pages they can fetch, understand and trust enough to repeat. This guide separates the work that official documentation supports from the tactics that are still experimental.

The names in circulation are SEO (search engine optimisation), AEO (answer engine optimisation) and GEO (generative engine optimisation). SEO is the established discipline. AEO and GEO are useful labels for the extra care of writing so an answer engine can lift a passage without guessing. They are not a replacement for SEO, and they are not a promise of inclusion.

How AI search and traditional search differ

A traditional results page is a list of links. The system ranks URLs and the person chooses one. An AI answer is a written response that may cite a few of those sources, or none. The person may never click.

That changes the shape of a good page more than it changes the foundations.

Google’s documentation on AI features and your website says the best practices for SEO remain the relevant ones, and that meeting the technical requirements still does not mean Google will crawl, index or show the page. Bing’s webmaster guidelines say the same crawling and indexing foundation supports Copilot and grounding, and that GEO does not guarantee citations.

ChatGPT, Gemini and Perplexity do not publish a stable public recipe for which business they will name. Treat any vendor that sells a guaranteed mention as selling something the platforms have not promised.

How assistants discover and cite pages

Public descriptions of these products point to three broad sources, which are not equally available to every query:

1. A search index. Google AI Overviews and AI Mode draw on pages eligible for Google Search. Bing and Copilot draw on Bing’s index. Some other assistants use a search partner rather than crawling your site themselves.
2. A live fetch. Some products crawl with their own bots, or fetch a page at answer time. If the bot is blocked, or the useful text only appears after JavaScript runs, that fetch can fail.
3. Model memory and third-party pages. A model may repeat a description it learned earlier, or a description published about you on someone else’s site. You cannot edit that memory on demand. You can publish a clearer canonical account and earn independent mentions.

A citation, when it happens, is usually a passage that answered the question directly: a definition, a comparison, a step, a price condition, a location. Pages written only as slogans give the system very little to lift.

Why technical SEO still matters

If a page is not crawlable and indexable, it is not a candidate for Google’s AI features. The same class of faults hurts every other engine.

Check these before rewriting headlines:

• The important text is in the HTML response, not injected only after client-side JavaScript.
• robots.txt and any CDN rules allow the crawlers you actually want. A disallow meant for a staging folder is a common way to hide a whole section.
• The page returns a successful status, is not noindex, and has one canonical URL that matches the URL you want indexed.
• The XML sitemap lists canonical URLs and is submitted in Google Search Console and Bing Webmaster Tools.
• Internal links use real <a href> elements. Buttons that only navigate with JavaScript are weaker discovery paths.
• Titles, meta descriptions and headings describe the page that is actually there. Duplicate titles across service pages make both look interchangeable.
• The page is usable on a phone and does not hide the main content behind an interstitial.

Core Web Vitals and accessibility are not “AI tricks”. They are part of whether people, and Google’s page-experience systems, can use the page. Fix the failures that block crawling first. Then improve the writing.

A platform migration is a frequent moment when this work is lost. The companion guide on rebuilding a WordPress site without throwing away SEO covers redirects, canonicals and the checks to run before launch.

Writing that is specific enough to cite

Answer engines prefer pages that state the thing plainly. A useful service page or guide usually does five things:

• Names the offer in the first screen, in words a customer would use.
• Says who it is for, and who it is not for, when that avoids a bad-fit enquiry.
• Answers the obvious questions in short paragraphs that still make sense if quoted alone.
• Shows evidence you actually have: a method, a constraint, a real project, a named technology. Invented percentages are not evidence.
• States limits. “We cannot guarantee a ranking” is more trustworthy than a promise, and it stops the page claiming something false.

Original experience matters more than volume. Google’s spam policies cover scaled content that adds little value, including mass pages produced with generative tools and published without review. Use AI to draft if you want. A person who knows the subject still has to check the facts, the examples and the claims before anything goes live.

For a business, the highest-value pages are usually the ones a buyer would ask an assistant about: the service, the locations you really serve, the migration or rebuild approach, and a small number of deep guides. Ten thin city pages will not outperform one accurate page.

Structure that is easy to extract

Structure is not a ranking hack. It is how a person, a screen reader and a retrieval system find the part that answers the question.

• One H1 that states the topic.
• H2s that are questions or concrete claims, not “Overview” and “Conclusion”.
• Short paragraphs. One idea each.
• A table when you are comparing options. A list when the order matters.
• A direct answer near the top, then the detail.
• FAQ answers that stand alone. Do not write “as above” or “click here”.

Write the heading as the question a customer asks, and make the next paragraph the answer. That pattern helps human readers whether or not an assistant ever quotes it.

Structured data, and the limit of it

Structured data is a machine-readable copy of facts that are already visible. Google recommends JSON-LD, and its structured data guidelines say that valid markup makes a feature eligible. It does not guarantee the feature will show. Markup that does not match the visible page, or that invents reviews, ratings or addresses, can make the page ineligible and can breach the guidelines.

Use it where it describes the page:

• Organization or ProfessionalService for the company, with the address you actually operate from.
• Service for a real service, with a description that matches the page.
• Article or BlogPosting for an article, with a real author and real dates.
• FAQPage only for questions and answers that are visible. Do not add it because you hope for a rich result. Google has narrowed when FAQ rich results appear, and an assistant is not obliged to quote the markup.
• BreadcrumbList when the breadcrumb trail is visible.

Do not add LocalBusiness for a city where you have no premises. Do not mark up star ratings you do not have. Bing’s guidelines say structured data may support clearer grounding and does not guarantee visibility.

On client work we mark up what the template already shows: organisation, services, articles, breadcrumbs and FAQs. That is hygiene. It is not a switch for AI inclusion.

Expertise, case studies and mentions elsewhere

Assistants repeat what they can corroborate. A single page on your own site saying you are the leading agency is a weak source. Stronger signals, when they are true, are:

• A named author with a real role and a consistent biography.
• Case studies that describe the work, the constraints and the outcome you can actually stand behind. AAMC Abogados is a WordPress-to-modern-platform rebuild. The public write-up does not invent ranking lifts, and this guide will not either.
• Independent mentions: a client’s site linking to you, a talk, a directory you genuinely qualify for, a journalist citing you. Buying mentions is a bad trade.
• Consistent names, URLs and descriptions across your site, your LinkedIn company page and the profiles you control.

If a number is not in the case study, do not put it in a guide “as an illustration”. Readers and answer engines both treat examples as claims.

Crawler access, without the folklore

robots.txt controls fetching. It is not a vote of confidence, and naming a bot with Allow: / does nothing extra if User-agent: * already allows the path.

Separate the decisions:

• Googlebot fetches pages for Google Search, which includes eligibility for AI Overviews and AI Mode. Blocking Googlebot removes you from Google Search.
• Google-Extended is a different token, used for some Gemini training and related controls. Google’s AI-features documentation does not say you must allow it in order to appear in AI Overviews.
• Bingbot fetches pages for Bing and, by extension, Copilot grounding.
• OAI-SearchBot, GPTBot, PerplexityBot, ClaudeBot and others are third-party crawlers. Allowing them lets those companies fetch the page. It is not a ranking factor in Google. Blocking them means that product cannot fetch you, which matters if its answers depend on a fresh crawl.

noindex keeps a URL out of Google’s index. Bing also documents controls such as NOARCHIVE and NOCACHE that can limit how much of a page Copilot may use. Read the current Bing guidelines before copying a robots snippet from a blog, including ours. The right policy depends on whether you want to be fetched.

llms.txt is a proposed Markdown file that offers a short, curated map of a site. It is reasonable to publish one if you will keep it accurate, as we do, because a contents page is useful to anything that bothers to request it. It is not a ranking factor. Google has said Google Search does not use AI text files of this kind, and that creating one will not help or harm Google rankings. Do not pay for an llms.txt as if it were the optimisation.

More than one language

If you serve people in more than one language, give each language its own URL, translate the page properly, and connect the versions with hreflang. Do not paste the English article into a Spanish URL. An assistant answering in Spanish will look for a Spanish source, and a duplicated English page is a poor one.

Translate the title, the meta description, the headings and the body. Keep the same facts. Localise examples where the market is different, rather than swapping a city name through one template.

GoNow’s own site is English and Spanish. Some client sites also include Catalan, because those audiences need it. The language set should follow the audience, not a checklist.

How to measure AI search, honestly

There is no single “AI visibility score” that tells you whether every assistant will recommend the business. Be wary of any dashboard that implies there is.

Measure what you can actually observe:

• Indexing and queries. Google Search Console for Google. Bing Webmaster Tools for Bing. Confirm the important URLs are indexed and which queries already show the site.
• Copilot citations, where Bing reports them. Bing Webmaster Tools has an AI Performance report for how Microsoft uses pages in AI answers. That report is about Microsoft’s experiences. It is not a ChatGPT report.
• Referral traffic. In analytics, look for referrals from AI products that send them. Many answers produce no click at all, so this number undercounts discovery. Treat it as a sample, not the whole story.
• Manual prompts. Keep a fixed list of questions a buyer would ask, run them periodically, and record whether you are named, whether a page is linked, and whether the description is accurate. Note the date and the product. This is qualitative research. A different account, region or phrasing can change the answer.
• Brand accuracy. When an assistant describes you wrongly, fix the page that should be the source of truth and any profiles you control. You cannot file a ticket to edit the model.

Report trends and corrections. Do not report a percentage lift you cannot reproduce.

Misconceptions that waste budget

• “Schema gets you into the answer.” It can clarify a page. It does not reserve a place in an answer.
• “llms.txt is the new sitemap, and it ranks you.” A sitemap helps discovery by search engines. llms.txt is optional and, for Google Search, unused.
• “Allowing GPTBot improves Google rankings.” These are different systems.
• “FAQ schema is a citation magnet.” Visible, self-contained answers help readers. The markup is not a promise.
• “We should publish a page for every city and every service.” Only where you have something distinct and true to say. Doorway pages are a spam pattern, not a strategy.
• “AI search has replaced SEO.” Google and Bing both say the opposite. The index is still the foundation.
• “A score out of 100 tells you if ChatGPT will recommend you.” It does not.

A practical checklist

Work through this in order. Stop when the next item is not the constraint.

• Confirm the money pages return 200, are canonical, and are not blocked or noindex.
• View the page source and check the main explanation is present without running JavaScript.
• Give each important page one clear H1, a specific title and a meta description that matches the page.
• Add internal links from related services, guides and case studies, with anchors that say what the destination is.
• Rewrite thin pages so the first paragraph answers what the page is, who it is for and what you will not promise.
• Add structured data that copies visible facts, and remove markup for reviews, awards or addresses you do not have.
• Connect translated URLs with hreflang only when the translation exists.
• Decide, in writing, which non-Google crawlers may fetch the site. Review that decision when a product changes its bot name.
• Publish or update llms.txt only as a maintained contents file, not as a ranking project.
• Set up Search Console and Bing Webmaster Tools, including Bing’s AI Performance report if you care about Copilot.
• Keep a short prompt log for the questions that actually generate leads. Review it monthly.
• When a claim is wrong or out of date, correct the page. Assistants repeat stale pages for a long time.

If you want this done against your own site, start with the SEO and AI search service. We work with businesses in Barcelona and, from Barcelona, with businesses in London. The assessment uses the same contact form as every other project. There is no automated score behind the button.

Sources

Checked against public documentation in November 2026:

• AI features and your website, Google Search Central.
• General structured data guidelines, Google Search Central.
• Top ways to succeed in AI search, Google Search Central Blog.
• Bing webmaster guidelines, including the sections on grounding, GEO and structured data.
• AI Performance in Bing Webmaster Tools.

Platform behaviour changes. If a statement here conflicts with the current page at those links, the official page wins. Third-party statistics about the percentage of searches that show an AI answer are not included, because they move quickly and are easy to overstate.