SEO for AI search: the 2026 checklist for Google, ChatGPT, Gemini and Perplexity

A company can rank third on Google for an important search and never appear in the answer ChatGPT gives to the same question. The reverse happens too: sites with modest rankings that are frequently cited by Perplexity or AI Mode. The reason is that ranking and citation are two different outcomes, produced by different systems. SEO for AI search is about working on both with the same website and the same content production.

The short answer is this: for an answer engine to cite you, it has to be able to read your pages (crawlers allowed and content indexed), it has to find passages in them that answer a specific question on their own, and it has to consider you a credible source on the subject. The rest of this article is the checklist we use to check this, divided into four areas: technical, content, authority and measurement.

For the general picture (what GEO is, who risks losing traffic) there’s our guide to generative engine optimisation. Here we go straight to the checks.

What’s the difference between ranking and being cited?

Ranking means appearing in a list of results ordered by relevance; being cited means being one of the few sources from which the AI takes a piece of its answer. In the first case the whole page competes; in the second, a passage does, often a paragraph or a table row.

The practical consequences:

• Few sources get cited, usually a handful compared with the ten organic links on a results page.
• The question is longer. AI Mode uses query fan-out: as Google explains, it breaks the question down into subtopics and runs many searches in parallel, so a page can come into play for just one aspect.
• A click isn’t guaranteed. A citation brings visibility even without a visit, and needs measuring with its own tools.

For AI Overviews and AI Mode, the link with SEO remains strong. In its guide to optimising for generative AI features, updated on 10 July 2026, Google writes that these features are based on Search’s ranking and quality systems, and that a page must be indexed and eligible to show a snippet in order to appear.

How does each engine choose its sources?

Each engine draws on a different index and feeds it with its own crawlers: knowing which bot reads your site on behalf of which service is the first check to make. Blocking the wrong bot in robots.txt is enough to vanish from an engine without noticing.

Engine	Where it gets its sources	Bots to let through	What to check	Where to measure
Google AI Overviews and AI Mode	Google’s index, with query fan-out	Googlebot	Indexing, snippet eligibility, no nosnippet on pages you want cited	Search Console, generative AI features report
Gemini (app)	Model plus grounding on Google’s index	Googlebot; Google-Extended governs training and grounding in the Gemini apps	Decide deliberately whether to block Google-Extended	Referrals from gemini.google.com in GA4
ChatGPT search	OpenAI’s crawler, plus other sources OpenAI doesn’t specify	OAI-SearchBot (GPTBot only concerns training)	Robots.txt, indexing on Bing as a precaution	Referrals with utm_source=chatgpt.com
Microsoft Copilot and Bing AI answers	Bing’s index	Bingbot	Bing Webmaster Tools, sitemap, IndexNow	Bing Webmaster Tools, AI Performance report
Perplexity	Its own index	PerplexityBot	Robots.txt and firewalls that don’t block it	Referrals from perplexity.ai in GA4

A few details that make the difference in projects.

Google-Extended has nothing to do with AI Overviews. According to Google’s crawler documentation, the Google-Extended token decides whether content can be used to train Gemini models and for grounding in the Gemini apps and on Vertex AI, and doesn’t affect inclusion or ranking in Google Search. Blocking it won’t remove you from AI Overviews and AI Mode, but it can reduce your presence as a source in the Gemini app.

For ChatGPT, OAI-SearchBot is what counts. On the page about its crawlers, OpenAI distinguishes three main bots: OAI-SearchBot for making sites appear in ChatGPT search, GPTBot for model training, and ChatGPT-User for actions initiated by a user (for which robots.txt may not apply). If you block OAI-SearchBot you won’t appear in ChatGPT search answers, except as a plain navigation link. OpenAI says it takes about 24 hours for a robots.txt change to be picked up. You can block GPTBot and leave OAI-SearchBot open, avoiding training while remaining citable.

Bing is still good insurance. OpenAI doesn’t publish how far ChatGPT search relies on external indexes beyond its own crawler, so having your site well indexed on Bing too is a cheap precaution. And it covers Copilot anyway, which is widely used in companies working with Microsoft 365.

Perplexity has two bots. According to Perplexity’s documentation, PerplexityBot indexes and respects robots.txt; Perplexity-User fetches pages at the user’s request and generally ignores it. To get into the index, the first is what counts.

Technical checklist: can the engines read your site?

The technical side is the quickest to fix and the one where a mistake costs the most, because a blocked bot wipes out your chances of being cited on that engine. Before touching the copy, it’s worth working through this list.

• robots.txt lets Googlebot, Bingbot, OAI-SearchBot and PerplexityBot through to the sections you want cited.
• You’ve made an explicit decision on Google-Extended and GPTBot, knowing what they govern.
• Firewalls, CDNs and anti-bot protection (Cloudflare and similar) don’t block AI search crawlers: check the server logs.
• Key pages are indexed on Google and on Bing (Search Console and Bing Webmaster Tools, with an XML sitemap submitted to both).
• Pages you want cited have no nosnippet, overly low max-snippet or data-nosnippet on important blocks. Google lists the same checks for AI features too.
• The main content is in the HTML text, not locked inside images, scanned PDFs or JavaScript that crawlers don’t run (check with Search Console’s URL inspection).
• A single H1, with H2s and H3s in logical order, descriptive internal links, and important pages reachable in a few clicks.
• Fast, stable pages on mobile (see Core Web Vitals).
• Valid JSON-LD structured data consistent with the visible text (more on this below).

llms.txt: is it worth it?

llms.txt is a proposed Markdown file, published at the root of the site, that summarises the main pages for language models. Google Search doesn’t use it: in its guide to optimising for generative AI, Google writes that to appear in Search, AI features included, you don’t need machine-readable files, AI text files or dedicated markup, because Search doesn’t use them. It adds that there’s no problem creating and maintaining an llms.txt file for other services that use it, and that the file has no positive or negative effect on the site’s visibility. None of the major answer engines has publicly stated that it uses it to choose sources.

Our approach: if the site can generate it automatically (we do so on digitalrunner.it), preparing it costs little and it may be useful to agents and tools that read it. But it comes after robots.txt, indexing and content. If someone offers it to you as the main lever for “getting into ChatGPT”, ask them what documentation they’re basing that on.

Does schema markup help you get cited?

Schema helps engines understand who wrote what and what kind of page it is, but on its own it won’t get a page cited. Google is explicit: in its documentation on AI features it says that AI Overviews and AI Mode need no additional optimisation or special schema, and in its optimisation guide it recommends continuing to use structured data as part of your overall SEO strategy, because it’s needed for rich results.

In practice we work on a small set: Organization (with sameAs pointing to official profiles), Person for authors, Article with publication and update dates, BreadcrumbList, and where it makes sense Product, Service, LocalBusiness, Event. The markup must match what the user sees: reviews or questions that don’t appear in the text are a risk.

Content checklist: are your pages citable?

Content is citable when it contains passages that answer a question completely even when read out of context, with a verifiable figure or statement. This is the part that takes the most work and the one that separates cited pages from pages that are merely indexed.

• Each H2 opens with one or two sentences that answer the section’s question. Detail comes afterwards.
• Some of the headings are phrased as the user’s real question, taken from related searches, “People also ask” and questions that have reached your salespeople.
• Each paragraph has one idea and works when extracted. Avoid vague references (“as mentioned above”, “this approach”) in the sentence that carries the answer.
• Comparisons go in tables, procedures in numbered lists, definitions in a single clear sentence.
• Every important page contains at least one element that can’t be found elsewhere: data you collected, an example from a project, a reasoned selection criterion, an original photo or video.
• Factual claims have a source, with a link to the primary source.
• A visible update date, and key pages reviewed when something changes (prices, regulations, platform features).
• No content mass-produced with AI without human input. In its guide, Google warns against “commodity” content based on common knowledge that could come from anyone, and against what a generative AI model could easily produce.

On the citability of passages, one clarification: Google writes that there’s no need to break content into small chunks for AI to understand it better. The goal remains a page well written for a person, with clear answers at the start of each section.

Topic clusters and hub pages

A single isolated article is rarely enough to become a recurring source. A group of connected content works better: a hub page covering the topic as a whole and a series of in-depth pieces, one for each sub-question, linked in both directions with descriptive internal links.

The reason is fan-out. Take a company that makes industrial filtration systems. A buyer asks AI Mode “what filter do I need for the dust from a paint line, how much does maintenance cost and what regulations apply”. The engine searches separately for filter types, maintenance costs and emissions regulations. A site with a hub page and three well-made in-depth pieces has three or four chances of being cited in the same answer; one with a single product page has one.

To build the cluster:

1. start from customers’ questions and the search intent of the main searches;
2. choose one hub page per topic (often a service page or a guide);
3. write an in-depth piece for each sub-question with commercial value and link it to the hub with descriptive anchors;
4. every six months, compare AI answers with your content and look for uncovered sub-questions.

Authority checklist: do the engines consider you a credible source?

Between two pages that answer well, engines tend to cite the one from a recognisable source, with identifiable authors and mentions elsewhere. Authority is built on and off your website, and it’s the part with the longest timescales.

• Every piece of content is signed by a real author, with an author page describing their experience and role and linking to their professional profiles.
• About us, contact details, company information and the description of the business are the same on your website, Google Business Profile, LinkedIn and industry directories. Inconsistent data confuses engines when they describe the brand.
• The brand is mentioned in publications, associations and industry sites (even without a link), through contributions, published data, events and interviews.
• You publish original material others can cite (a customer survey, a benchmark, a technical guide) and you’re present where engines often look, such as YouTube and industry communities.

Avoid manufactured mentions: Google’s guide says that seeking inauthentic mentions across the web isn’t as useful as it might seem. One genuine citation in a trade magazine is better than twenty bought articles.

Measurement checklist: how can you tell whether you’re being cited?

As of October 2026 you can measure three things: impressions in Google’s AI features, citations in Microsoft services and traffic arriving from assistants. The rest (presence in ChatGPT or Perplexity answers) requires manual checks or third-party monitoring tools.

• Search Console, generative AI features report. Google announced it on 3 June 2026 and since 31 August 2026 it has been available for every site. It shows impressions, pages, countries, devices and trends over time for AI Overviews and AI Mode, with a separate report for AI features in Discover. Clicks, CTR and queries are missing: you know which pages appear, not for which questions. The same impressions also remain in the general performance report.
• Bing Webmaster Tools, AI Performance report. In public preview since 10 February 2026, it shows how often your pages are cited in Copilot, in Bing’s AI answers and in some partner integrations, with grounding queries and a page-by-page breakdown.
• GA4, a channel group for AI assistants. Create a custom channel that collects referrals from chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com and the like. ChatGPT adds utm_source=chatgpt.com to its links, so you’ll also find it among your sources.
• Monthly manual checks. A fixed list of 20-30 questions from your market, put every month to AI Overviews, AI Mode, ChatGPT, Gemini and Perplexity. Note whether you’re cited, who appears in your place and how the brand is described.
• Monitoring tools. Ahrefs Brand Radar, Semrush or Profound measure your share of citations across a set of prompts compared with competitors. Useful for the trend, bearing in mind that they sample answers that vary. Be wary of anyone boasting of “internal” Google data: Google itself warns that no third-party tool has access to its ranking or AI systems.

The Search Console report on AI features, in the form announced, focuses on impressions: to understand their value, it needs cross-referencing with organic traffic and conversions.

What changes for lead generation and the CRM

For a B2B company, a citation is worth something if it brings qualified enquiries, and those enquiries need to be recognised as such in the CRM. People arriving from an AI answer have often already compared suppliers and options: fewer visits, but closer to a decision.

Three measures we apply in our projects:

1. Track the source all the way to the opportunity. The AI assistant referral must reach the CRM as the contact’s source, together with the landing page; otherwise the “organic” bucket contains everything. We discuss this in our article on metrics for managing B2B leads.
2. Ask in the form how they found you. People who discover you on ChatGPT often come back by searching for the brand on Google, and show up as branded or direct traffic. A “How did you hear about us?” field with an “AI assistant” option recovers what analytics misses.
3. Prepare the landing pages. Cited pages are often articles or guides: they need a next step consistent with the question (a data sheet, a configurator, a call with an engineer), not just the generic form.

For AI Mode in particular, with its effect on leads and on the journey of people looking for a supplier, we’ve written a dedicated piece: how to get cited by Google AI Mode and turn citations into leads.

The mistakes we see most often

• Blocking AI crawlers wholesale. Rules copied from the internet that lump together GPTBot, OAI-SearchBot, PerplexityBot and Google-Extended. The aim was to avoid training; the result is disappearing from ChatGPT and Perplexity search.
• Relying on robots.txt alone. The file is correct, but the CDN’s firewall rejects the bots. You only find out from the logs.
• Ignoring Bing. A site never verified in Bing Webmaster Tools starts from zero on Copilot.
• Treating llms.txt as the solution. Weeks spent on a file Google says it doesn’t use, while key pages are left without clear answers or an author.
• Rewriting everything “for AI”. Text broken into fragments, repeated synonyms, bloated FAQs: Google’s guide says it isn’t necessary.
• Measuring only clicks. Impressions in AI features rise, clicks fall, and the conclusion is that “SEO doesn’t work any more”. The right figure is how many qualified leads arrive, and from where.

Want to know where you stand today?

Our AI visibility and SEO audit starts from this checklist applied to your website: what crawlers see, which questions you’re already cited for and where competitors appear, which pages are worth fixing first and how to connect it all to leads in the CRM. At the end we’ll tell you frankly whether it makes sense to invest in GEO now or whether there are more urgent priorities on the site.

You can book a call with Federico or write to us via the contact form.

Frequently asked questions

Do you need different SEO for ChatGPT and for Google?

The fundamentals are the same: indexable pages, useful content, a recognisable source. What changes is the crawlers to let through (OAI-SearchBot for ChatGPT, Googlebot for Google) and the indexes each engine draws on.

Will blocking Google-Extended remove me from AI Overviews?

No. According to Google’s documentation, Google-Extended governs the training of Gemini models and grounding in the Gemini apps and on Vertex AI, and doesn’t affect inclusion or ranking in Search. AI Overviews and AI Mode use Googlebot’s normal crawling.

Does llms.txt help you get cited by AI?

Not for Google: Search doesn’t use it. The other major engines haven’t stated that they use it to choose sources. You can publish one if it costs little, without making it a priority.

Can I block training and still be citable on ChatGPT?

Yes. You can block GPTBot, which OpenAI uses for training, and leave OAI-SearchBot open, which powers ChatGPT search. Changes to robots.txt are picked up in about 24 hours.

How can I see whether I appear in AI Overviews and AI Mode?

In Search Console, in the report dedicated to generative AI features, available for every site since 31 August 2026. It shows impressions, pages, countries and devices.

Does schema markup get a page cited?

Not on its own. Google says there’s no special markup for AI features; structured data helps engines understand author, page type and dates, and must be kept consistent with the visible text.