Generative Engine Optimization Checklist: 14 Pass/Fail Tests In Fix Order

A generative engine optimization checklist is only worth running if every item can fail. Test one live page for crawler access, an answer in the first third, sourced evidence, consistent entities and outside mentions, then fix in that order, because a page an AI crawler cannot reach never gets cited, however well the page reads.

This applies to any informational or comparison page on a B2B site that you want ChatGPT, Perplexity or Google AI Overviews to quote.

No checklist guarantees a citation. The engines choose differently by query, and Google says AI features need no special optimisation beyond standard Search eligibility.

How To Use This Checklist

Most GEO checklists are lists of tips. Tips cannot fail, so nobody knows when they are done. This one is 14 tests. Each has a yes or no answer, a way to run it and a fix, and they are grouped into six tiers in the order that does the most good first.

If you need the definition of GEO and how it differs from SEO and AEO, read AISO versus SEO, AEO and GEO before you start. This page assumes you already have it.

The checklist tests whether AI engines can parse, understand and reference one page, not whether that page ranks. Every item scores pass or fail, with no partial credit.

Run it like this.

1. Pick one live page that targets a single question a buyer asks. Not the homepage.
2. Score each of the 14 items as pass or fail, using the test written under it.
3. Write down every failure in the scorecard at the end of this page.
4. Fix the failures in tier order, Tier 1 first.
5. Rerun the same tests on the same page and compare with your first score.

It does not cover a site-wide technical audit, the definition of GEO or a roundup of AI search statistics. Where one of those matters, the relevant section links to the page that handles it.

Is GEO a separate discipline from SEO? In practice you are fixing one page for two audiences, a searcher and an engine that quotes the page. Most of the fixes below also make the page better for people, which is the quickest way to tell a useful item from a ritual.

Tier 1: Access Checks (Run These First)

An engine cannot quote a page it cannot read. These three items come first because a failure here makes every later fix irrelevant. Each takes a few minutes.

Item 1: The Search Crawlers You Care About Are Allowed In Robots.txt

OpenAI documents separate crawlers for search, training and user visits, each controlled in robots.txt, and blocking the search crawler removes a site from ChatGPT search. That makes robots.txt the first file to read, because one blanket rule can switch off a whole engine without anyone noticing.

• Test: Each crawler named in the vendor documentation for the engines you want to appear in is allowed to fetch the page.
• How to check: Open yourdomain.com/robots.txt. For each engine, find the crawler names in that vendor's own documentation, then look for any Disallow rule, including a wildcard rule, that covers the page's path.
• Fix: Remove or narrow the rule that blocks the search crawler. Decide training and user-visit crawlers separately, because each has its own control.
• Pass condition: No Disallow rule covers the page for any search crawler you want to be cited by.

Item 2: The Answer Is In The Delivered HTML

If the opening answer only appears after a script runs or after a sign-in, you are betting that every engine renders pages the way your browser does. You do not need to make that bet.

• Test: The opening answer is present in the HTML returned without running script and without signing in.
• How to check: View the page source, or fetch the URL with script switched off, and search for the first sentence of your answer.
• Fix: Move the answer into server-rendered HTML, or remove the login or interstitial that sits in front of it.
• Pass condition: The first sentence of the answer is found in the raw source.

Item 3: The Page Is Indexable And Eligible For Standard Search

Google says AI features need no special optimisation, only standard Search eligibility. So the boring checks are the real checks.

• Test: The page returns a 200 status, carries no noindex directive and has a canonical tag that points to itself.
• How to check: Use your browser's network panel for the status, search the source for a noindex robots tag, and read the canonical tag.
• Fix: Remove the noindex, fix the redirect chain, or correct the canonical so it names the page's own address.
• Pass condition: All three conditions hold at once.

Failures across many pages at once are a site problem, not a page problem. That is where Technical SEO and AI readiness work earns its fee, because a person has to read your crawler rules and your rendering setup together.

Tier 2: Structure Checks (Is The Answer Easy To Lift?)

Once a crawler can reach the page, the next question is whether an engine can pull a clean answer out of it. These three items test the shape of the page. For fuller examples, see how to structure AI-era content.

Item 4: A Direct Answer Sits In The First 30 Percent Of The Page

Position matters because engines appear to lift from the top. In a CXL study of 100 AI Overview citations, 55% of the cited passages came from the first 30% of the source page.

• Test: The first direct answer to the page's main question appears within the first 30% of the page.
• How to check: Find the first sentence that answers the question the page is built around. Divide its position by the total length of the page.
• Fix: Move the answer up and cut the scene-setting above it. Put background below the answer, not before it.
• Pass condition: The answer starts before the 30% mark.

Item 5: Each H2 Opens With Its Own Answer That Works Alone

Engines quote passages, not pages. A section that opens with "As mentioned above" cannot be quoted without the part before it.

• Test: Each H2 opens with a 40 to 60 word answer that still makes sense with everything around it removed.
• How to check: Copy the opening paragraph of one H2 into a blank document. Read it cold. If you need the surrounding text to understand it, it fails.
• Fix: Replace pronouns with the nouns they stand for, attach the qualifier to the claim in the same sentence, and put units and periods on every number.
• Pass condition: Every H2 opening passes the cold read.

Here is the difference on a single sentence.

• Fails the cold read: It also costs more in the first year, so teams usually skip it.
• Passes the cold read: A dedicated rank tracker costs more in its first year than a spreadsheet, so small teams often start with the spreadsheet.

Item 6: Comparisons And Steps Live In Tables And Lists

A comparison buried in a paragraph is hard for a reader to scan and harder for an engine to extract.

• Test: Any comparison of three or more things is a table, and any sequence of steps is a numbered list.
• How to check: Scan the page for sentences that line up three or more options in a row, or that say "first, then, finally".
• Fix: Rebuild the passage as a table with the same attributes for every row, or as an ordered list.
• Pass condition: No comparison of three or more things is left in running prose.

Tier 3: Evidence Checks (Can A Claim Be Traced?)

An engine choosing what to quote favours claims it can attribute. These items test whether a reader, or a model, can trace each claim to a named source. For a worked programme on a live site, see our GEO case study.

Item 7: Every H2 Carries A Sourced Statistic Or Finding

A workable pass condition is at least one sourced statistic or finding per H2 section, each linked to a named source in the same sentence. The research behind that habit is early. A 2023 Princeton-led preprint found that adding statistics, citations or quotations raised visibility by up to about 40% in a simulated generative engine rather than a live product, while keyword stuffing gave minimal gain.

Read that as a reason to try the habit, not as a promise of a 40% lift on your page.

• Test: Each H2 section contains at least one statistic or finding with a named source linked inline.
• How to check: Go section by section and count. A section with no sourced number or finding fails.
• Fix: Add a figure you have actually checked, linked on a short phrase in the sentence. If you cannot find a real source, cut the claim, because an unchecked number is worse than none.
• Pass condition: Every H2 has at least one.

Item 8: The Page Shows A Date And No Figure Has Gone Stale

An old number on a page that looks fresh is a quiet way to lose trust.

• Test: The page shows a visible published or updated date, and every figure on it matches its source today.
• How to check: List each number on the page, open its source, and compare both the figure and the period it covers.
• Fix: Update the figure and the date together. If the source has moved or vanished, remove the number.
• Pass condition: The date is visible and no figure disagrees with its source.

Item 9: Quotes And Findings Name Their Source

"Studies show" and "experts say" give an engine nothing to attribute and give a reader nothing to check.

• Test: Every quote or finding is attributed to a named person, organisation or document.
• How to check: Search the page for "studies show", "research suggests" and "experts say".
• Fix: Name the source and link it on a short phrase, or delete the sentence.
• Pass condition: No unnamed authority remains on the page.

Tier 4: Entity Checks (Does The Page Say Who And What?)

Engines build answers from named things. If your page calls the same company three different names, or leaves its subject as "it", the engine has to guess. Two items test this. For tooling that checks entity coverage, see these entity-first on-page tools.

Item 10: The Author And Company Are Named Consistently

• Test: The page names its author and the author's role, and the company is written the same way as on the About page.
• How to check: Compare the byline and company name on this page against two other pages and the About page.
• Fix: Pick one form of the company name and one form of each author's name and role, then apply them everywhere.
• Pass condition: The name, role and company match across all pages compared.

Item 11: Every Concept Is Named In Full On First Use

A page that names its entities in full on first use is easier for an engine to attribute. The rule is simple and the failures are everywhere.

• Test: Every product, standard or concept the page relies on is named in full the first time, and no section opens with a bare "it" or "this".
• How to check: Read only the first sentence of each section. Any pronoun with nothing before it fails.
• Fix: Replace the pronoun with the noun, and spell out the full name at first mention.
• Pass condition: No section opening depends on the section before it.

The difference looks like this.

• Wrong: This also helps with the crawl, so it should be done early.
• Right: Allowing the search crawler in robots.txt also helps the crawl, so the robots.txt check should be done early.

Entity consistency across a whole site is a bigger job than one page. Our LLM SEO work covers that.

Tier 5: Earned Mention Checks (Are You Named Elsewhere?)

Everything so far is on your own page. This tier looks outside it, and it is the one most checklists skip.

Item 12: You Appear On A Source The Engine Already Cites

The sources an engine cites for a query are visible in its answer and can be read directly. That makes this test cheap and honest. It describes where you stand. It does not claim that a mention causes a citation.

• Test: The brand or page appears on at least one third-party source that the target engine already cites for the query, or you can name a realistic way onto one.
• How to check: First, run the query in the engine and open every source it cites. Second, check whether your brand or page appears on each, and note the ones where it could.
• Fix: Pitch a contribution, a correction or a data point to the cited publications that fit. Our earned mentions service does this work for clients, and our own AEO case study shows the approach applied to our own site.
• Pass condition: You appear on at least one cited source for the query, or you have a named target list for the sources you do not.

Tier 6: Measurement Checks (Did The Fix Work?)

Without a baseline, every improvement is a story. These two items make the work provable.

Item 13: A Dated Baseline Of 10 To 20 Queries Exists Before You Change Anything

Citation testing is commonly done by running 10 to 20 relevant queries on each engine each month. Write the list before you touch the page, and record the sources each engine cites.

• Test: A written list of 10 to 20 queries exists, with the cited sources recorded per query and per engine, and a date on it.
• How to check: Open the document. If it is missing, undated or has no sources recorded, it fails.
• Fix: Build it now, before any fix goes live. Use the table below.
• Pass condition: The list exists, is dated, and was written before the changes.

Repeat the three rows for every query on your list.

We keep the baseline query list and the ranking tracking for the same pages in Semrush, so the AI answer checks sit next to the Google positions for the same queries. If you want software built for the reruns, start with this guide to answer engine optimization tools.

Item 14: The Same Queries Are Rerun On A Schedule

• Test: A second run of the same queries on the same engines exists, dated, and compared with the baseline.
• How to check: Open the baseline and look for a later run next to it.
• Fix: Put the rerun in the calendar. Change nothing about the wording of the queries between runs.
• Pass condition: At least one dated rerun sits beside the baseline, with the differences noted.

Where the results go once you have them matters too. A monthly AI visibility report template gives the baseline and the reruns a home your stakeholders will actually read.

What You Do Not Need To Add

Many published checklists score an llms.txt file, FAQ schema and speakable schema as pass or fail items. Google's documentation says AI features need no special optimisation, no new AI-specific text files and no special schema markup.

So we mark all three as optional, not as failures. A page does not fail this checklist because it lacks them.

FAQ sections themselves are still useful for readers who arrive with a follow-up question. Keep them for that reason. Just do not add markup to a page because a checklist said it was mandatory.

The Fix Order: What To Repair First

Access failures block everything, so they come first. Answer placement comes next, because the CXL finding on position is the strongest evidence in this checklist. Sourced evidence follows. Entities and outside mentions come after that, and measurement runs alongside all of it.

The first three fixes in order are the crawler rules, then the answer in the first 30% of the page, then a sourced statistic in each section. The crawler comes first because a blocked search crawler removes a page from ChatGPT search. The answer comes second because of the position finding. The statistic comes third because it is the habit with the most research behind it, which is still a simulated study.

The research supports the first two placements for access and answer position. Everything below that in the table is a judgement call from our own practice, not a measured ranking.

Technical items outside a single page are covered in AI visibility audit versus SEO audit. Once the page-level fixes are done, the next step is the AI search visibility strategy that decides which pages to fix next.

Common Mistakes When Running A GEO Checklist

• Scoring partial credit: A page that half passes still fails. Rewrite the item until it is a clean yes or no.
• Auditing the homepage: A homepage rarely targets a question. Run the checklist on a page built to answer one.
• Adding schema because a checklist said so: Markup does not replace a page that fails the tests above. Fix the page first and add markup only where it helps a reader or a rich result you actually want.
• Citing a vendor statistic with no primary source: If you cannot find where the number came from, cut it.
• Rerunning tests without a baseline: A second run has nothing to compare against. Write the baseline first, then fix, then rerun.

Worked Example: Scoring One Page

Copy the scorecard below for your own page. Mark each row pass or fail, count the passes, then brief the first three failures in tier order.

Here is the arithmetic with illustrative numbers you replace with your own. Say your page passes 9 of the 14 items. That is a score of 9 out of 14, which leaves 5 failures. Take the three lowest-numbered failures, because the lowest numbers sit in the earliest tiers, and write one brief for each.

A brief needs four lines.

• Item: The number and name of the failed test.
• Evidence: What you saw when you ran it.
• Fix: The change to make, in one sentence.
• Owner and date: Who does it and when it is rechecked.

If a page fails so many items that patching is slower than rewriting, a Content Refresh is usually the faster route.