What Checklist Should I Use for AI Search Optimization?

The checklist to use for AI search optimization has two halves. The first half is eligibility: can ChatGPT, Perplexity, Gemini and Google's AI Overviews fetch, index and read your pages at all. The second half is visibility: once they can, is there a clear, sourced answer on your site and on the third-party pages they cite that makes them pick you. Each line below has a pass condition, an action, an owner and a recheck signal, and every rule is dated to the source it came from, because some of these rules have already changed this year.

I wrote it because the checklists I kept finding were either a strategy process with the operating detail left out, or a list of tactics with no way to tell whether a line passed. Both have their place. A marketing lead trying to assign work on a Monday needs something closer to a test sheet.

What AI search optimization is, and what it can't do

AI search optimization is the work of making your pages, and the pages other people publish about you, easy for AI answer engines to retrieve, read and quote, so that when a buyer asks a question in your category the answer names you and links to you. People also call it GEO or AEO. Google's position, in its guide to generative AI features in Search (updated 10 July 2026), is that AEO and GEO are SEO as far as Google is concerned.

The limits are worth stating before the checklist, because they decide what a "pass" can mean.

You control eligibility. You can make every priority page fetchable, indexed, rendered without JavaScript and canonical, and you can verify each of those with a request and a log line.

You only influence selection. A July 2026 survey of 45 GEO studies by Olivier Martinez found that content already retrieved into an engine's context can change how it gets cited, but that no reviewed technique showed a stable, long-running, cross-platform causal effect on being discovered in the first place. The same survey reports that citation-oriented rewrites can hurt retrieval. So the visibility half of this checklist is a set of conditions that make a citation more plausible, and you measure the outcome separately.

Nobody outside the engines can see their ranking signals. Google's guide says to be wary of tools that claim to use internal Google metrics. That applies to us too: CueScout measures what the answers say from the outside, by asking the questions and storing what comes back.

What the existing checklists already cover

The two pages I'd read first are both good, and this post is meant to sit under them, not replace them.

Aleyda Solis's AI search optimization checklist (updated 26 September 2026) is a 12-step process: build a prompt library, measure presence, diagnose the gap, fix readiness, entities, decision content, third-party sources, commercial data and localization, then report and rerun. It prioritizes gaps with Severity x Importance / Effort and asks for an owner per gap. If you're designing the program, start there. Step 4 tells you to audit priority URLs for crawlability, indexability and rendering; it doesn't spell out what passing looks like for each, which is what most of this post does.

Google's guide is the only first-party rulebook for AI Overviews and AI Mode. It's useful mostly for what it rules out: no new machine-readable files, no AI text files, no special schema, no chunking content into small pieces, no ideal page length. It doesn't mention Google-Extended at all, and it says nothing about the other engines' crawlers, which is where most of the robots.txt mistakes I see come from.

Part 1: Eligibility checks

These four decide whether an engine can use your page. A failure here on a page you care about outranks everything in Part 2.

Crawler access: allow the search agents, decide training separately

Every engine now runs more than one agent, and robots.txt treats them separately. The split that matters, from each company's own documentation as of today:

OpenAI states that sites opted out of OAI-SearchBot won't be shown in ChatGPT search answers, and that disallowing GPTBot only signals that content shouldn't be used for training. Google states that Google-Extended doesn't affect inclusion in Google Search. So a site that blocked GPTBot and Google-Extended to stay out of training has lost nothing in search, and a site whose User-agent: * group disallows everything, with exceptions only for Googlebot and Bingbot, may have dropped out of ChatGPT, Claude and Perplexity search without anyone deciding to.

Pass condition: for each engine you want to appear in, its search agent is allowed on your priority paths, and a request with that user agent gets a 200 from your CDN, not a challenge page. Robots.txt allowing an agent is no use if a bot-management rule blocks it first, so check the server or CDN logs for the agent's actual requests.

Two more details. ChatGPT-User and Perplexity-User fetch pages when a person asks, and both companies say these fetchers may not follow robots.txt, so they aren't a reliable opt-out lever either way. And Perplexity says robots.txt changes can take up to 24 hours to apply; OpenAI says about the same for OAI-SearchBot.

Our own GEO audit reads robots.txt for GPTBot, ClaudeBot, PerplexityBot and Google-Extended. That covers Perplexity's search agent and the two training crawlers people block most often, but not OAI-SearchBot or Claude-SearchBot, so check those two lines by hand until we add them.

Indexability: the Google rule is the whole rule

For Google's AI features the eligibility rule is short. A page must be indexed and eligible to appear in Search with a snippet, and the guide adds that the site has to be included in Search generative AI features in Search Console. That's it. A nosnippet directive on the content you want quoted works against you, since Google's own condition is snippet eligibility.

Pass condition: each priority URL returns 200, has no noindex, shows as indexed in Search Console's URL Inspection, and carries no snippet restriction on the parts you want quoted. Action: fix whatever fails, then request indexing. Recheck after releases, since a template change is the usual way a noindex reaches production.

Rendering: the answer has to be in the HTML

Vercel and MERJ reported in December 2024 that none of the major AI crawlers they monitored rendered JavaScript, naming OpenAI's crawlers, ClaudeBot and PerplexityBot. Gemini was the exception because it uses Googlebot's rendering. That test is almost two years old and crawlers change, but it's the most specific public evidence I know of, and it costs nothing to design for.

Pass condition: fetch the page with curl or view-source and find the answer paragraph, the product name, the prices and the comparison table in the raw response. Action: server-render or pre-render that copy. A client-only shell reads as an empty page to a crawler that doesn't run your app.

Canonical pages: one page per question, one value per fact

Google's guide tells you to reduce duplicate content. The AI-specific reason is that engines quote numbers, and if your pricing page, an old launch post and a help article give three different prices, the engine picks one and you don't get to choose which.

Pass condition: each priority topic has one canonical URL with a self-referencing rel=canonical, and the facts on it (price, plan limits, supported integrations) match every other page that repeats them. Action: consolidate overlapping pages, redirect the losers, and put a "checked on" date beside any number that changes.

Part 2: Visibility work

Passing Part 1 makes you readable. These four make you worth quoting. None of them has a binary pass the way robots.txt does, so the pass conditions are about what you can observe on your own pages and in the answers.

Entity clarity: one description, repeated everywhere

When an assistant can't find a plain statement of what you are, it improvises, and the improvisation is often wrong. Aleyda's checklist recommends a dated, approved fact sheet of names, categories, audiences, plans and canonical profiles, compared against your site and the listings that describe you. I'd copy that exactly.

Pass condition: your homepage, about page, pricing page and your G2, LinkedIn or Crunchbase profiles agree on the product name, the category, who it's for and how it's priced. Organization schema with sameAs links to those profiles helps tie your name to an entity the engines already know. Google says no special schema is required for its AI features, so treat the markup as cheap insurance rather than a lever. The CueScout audit checks for Organization schema with sameAs, a reachable about page and a public pricing page, because those are the three places assistants go to describe you.

Extractable answers: match the format to the question

Our Claude citation study found the format of the cited page followed the format of the question: for "best X" questions, 77 of 86 citations were listicles; for "A vs B" questions, 86 of 99 were comparison pages; for how-to questions, 94 of 110 were how-to guides. That's one engine and 60 answers, but the pattern was clean enough to plan around. A pricing explainer won't win a comparison question, however well written.

Pass condition: each priority page states its answer in the first few sentences under the title, using the title's own words, with headings that name the questions they answer and tables where the content is a comparison. Action: rewrite the opening, then match the page type to the prompt type. Watch the overcorrection, though. Google explicitly says you don't need to break content into tiny pieces, and the Martinez survey found citation-oriented rewrites can impair retrieval. Clear sections are the goal; a page shredded into 40-word fragments isn't.

Evidence and sources: give the engine something to lift

The 2024 Princeton paper behind the term GEO found that adding statistics, quotations and cited sources raised a page's share of the generated answer, which I covered in the GEO guide. The Martinez survey adds the caveat that those gains were measured on sources already placed in the engine's context, so they say nothing about getting retrieved. Both can be true: specifics help once you're in the candidate set.

Pass condition: every claim on a priority page that a buyer would act on has a number, a date or a named source next to it, and prices carry a "checked on" date. Action: replace adjectives with figures you can defend and link the source. Google's guide asks for a unique point of view and first-hand experience, and data you collected yourself counts as both.

External source coverage: know which outside pages decide the answer

This is where AI search parts ways with a ranking report. In one customer audit we ran, 58 of 96 pages the engines cited didn't rank in Google's top 100 for the query that surfaced them (details here). And the engines don't read the same pages: on one marketing-automation question, five engines cited 44 pages from 27 sites, and no site was cited by all five (the 500-answer comparison).

In our 533-URL corpus study of the AI-visibility tools category, how many of the 30 most-cited pages a brand appeared on tracked its unbranded visibility closely: one brand was on 21 of 30 and had 30% visibility, another on 1 of 30 with 2%. That's one category and a correlation, but it's the reason this check is on the list.

Pass condition: for each priority prompt, you have the list of cited URLs per engine, you know whether you're on each one, and every page you're missing from has a decision: pitch it, correct it, or leave it. Action: collect the cited pages, then work the short list. Budget for it honestly, since roundup placements are frequently paid. AI citation tracking stores those URLs per question and engine and marks whether your brand is on each page.

How to order the work

Score each failing check on four things, then work from the top. The ratings below are my defaults for a typical B2B software site, not measurements. Your own evidence should overwrite them.

Two rules sit above the scores. An eligibility failure on a priority page goes first, because nothing in Part 2 can work on a page the engine can't read. And a decision-stage accuracy problem, such as a wrong price in an answer to a "which should I buy" prompt, beats a missing mention in a broad definitional answer, even if the definitional prompt gets asked more. I wrote the longer version of this, with hard gates and tie-breakers, in how to prioritize content tasks from AI visibility findings.

Rechecking without fooling yourself

Eligibility checks are deterministic. Rerun the same request after the fix and it either passes or it doesn't. Visibility doesn't work that way, because answers vary between runs. In our 500-answer data, the sites Perplexity cited overlapped 73% from one day to the next and ChatGPT's 23%. Aleyda suggests running each core prompt three to five times inside a fixed 24 to 72 hour window, which is a reasonable floor.

So keep two records per task. One says the work is done: the URL you shipped and the date. The other is the answers on the next sweeps of the same prompts. A later citation is an observation, and you won't know for certain that your change caused it. That separation is how To-dos work in CueScout: you mark a task done, audit tasks can be rechecked against the same page, and whether the answers moved is measured separately on the following sweeps.

If you want the eligibility half run for you, the GEO audit does eight site checks and five page checks with the fetched evidence attached, and the free AI Readiness Checker runs the same eight site checks on any URL without a signup. The visibility half needs your prompt set and someone with time, and no audit replaces that.