AI Search and AEO

Run a GEO content audit before you publish another article

 ·  12 min read

Source documents connect to an answer card beneath a magnifying glass
In this article

A GEO content audit checks whether a page answers a buyer’s question with evidence an AI search engine can use. Start with the question, test access, trace each claim to its source, then record citations and business results separately. Define how you will test each change before you publish it.

For a marketing director, that means a smaller, more useful work list. You might need to repair a weak comparison, explain a service boundary or replace an unsupported claim. Another broad article about your category may add little to the decision a buyer needs to make.

We recommend auditing one commercial page and one supporting article together. The commercial page explains the offer. The article helps the reader decide whether that offer fits. Keep the audit small enough that you can inspect every source and explain every change.

Choose the buyer question first

Write the question in the buyer’s words, then name the decision it supports. “What is automation?” covers a wide subject. “Can we route qualified leads into our CRM without replacing it?” points to an integration decision, a technical constraint and a likely service need.

Use actual questions from sales calls, site searches or support conversations where you have permission to use them. Remove names and private business details. If you start with a proposed question, label it as a hypothesis in your working notes. A plausible prompt does not establish search volume or buyer demand.

Next, check your existing pages. Search the site for the question and its close variants. Choose the page that already answers it most closely. If two pages make the same promise, decide which one should carry the full answer before adding a third. Give supporting pages a narrower purpose and link them to the main explanation.

For each candidate, write one sentence that completes this thought: “After reading this page, the buyer can decide whether…” If you can’t finish it without using a broad verb such as “understand,” narrow the task.

Our guide to GEO services covers the wider buying decision. Keep a page audit focused on the text, evidence and next step on one URL. That distinction also helps you avoid publishing several posts that compete for the same question.

Check that the engine can reach the answer

Test the exact page URL and the text that contains the answer. A working homepage tells you little about a blocked article, a login screen or a page that loads its main explanation only after someone clicks a button. Record what you actually tested.

Start with an ordinary browser visit, then inspect the page response and rendered text. Check its HTTP status, canonical URL, indexing controls and robots rules. Look for a redirect to a different page, a security challenge or an empty content area. Keep the response details beside the audit finding so the person fixing it can reproduce the problem.

OpenAI’s crawler documentation separates OAI-SearchBot, which serves search, from GPTBot, which collects material for model training. Their settings work independently. Audit the search crawler’s access rather than treating a training permission as a search requirement. Check firewall behavior as well as the robots.txt file.

A test request with a bot name proves only how your server handled that request. It doesn’t prove that the real crawler visited, indexed or selected the page. When you have server logs, check genuine requests against the provider’s verification guidance.

Our AI crawler access checker gives you a starting point. Follow any finding through to the affected page. Save the old setting before a change, then repeat the same test afterward. Keep “access works” separate from “the engine cites us” in your report.

An archivist follows a red thread from an open book back to a source folder, beside a path from a web doorway.

Trace every useful claim to its source

Inspect the sentences a buyer might use to choose a supplier, set a budget or approve a project. For each sentence, ask where the fact came from, when it applied and how much the source actually supports. Put the source beside the claim while you edit.

A link at the end of a long paragraph can leave the reader guessing. Move it next to the fact it supports. If a source covers one product tier, don’t extend its promise to every tier. If it reports a sample of businesses, keep the sample and measurement period attached to the result.

Use a small evidence record for each important claim:

  • The exact sentence on your page.
  • The source page or original record.
  • The date and scope of the evidence.
  • The limitation a reader needs to know.
  • The action: keep, narrow, update or remove.

Treat a supplier’s documentation as evidence for its own stated features. Treat your own test as evidence for the setup you tested. Neither automatically proves what every customer will experience. When you describe an estimate, show its inputs and calculation so the reader can change an assumption.

Check the destination itself. A citation that leads to an error page or an unrelated report cannot support the sentence. If the original source disappears, look for a legitimate archived copy or another primary source. Don’t replace it with a second article that repeats the same claim without evidence.

Keep an explicit “unknown” in your working notes when you can’t verify something. In the public copy, remove the claim or explain the missing information if it affects the reader’s decision. Filling that gap with a confident number makes the audit less useful.

Read GitHub findings within their limits

GitHub can show the code, data and assumptions behind a GEO recommendation. Read those parts before adopting a headline. Check whether the work tests real search results, a controlled answer system or one publisher’s traffic. Those designs answer different questions.

The GEO research repository reports visibility gains of up to 40% in its experiments. That result supports further testing of content changes. It doesn’t promise a 40% traffic increase for your site, or a fixed position in a commercial assistant’s answer.

The get-cited-by-ai project encourages checking demand before prescribing page changes. Its stated limits include six owned sites, observational evidence and a strong Bing Copilot focus. Use its findings to form a question you can test in your market; keep those limits attached.

The Glasp AEO experiment compares changed pages with other pages on the same domain. Its authors report a placebo-test p-value of 0.16 and describe the evidence as suggestive. Compare outcomes before crediting a content edit for all subsequent growth.

For each recommendation you keep, record a prediction. For example: “Adding the supported integration boundary should help readers judge fit.” Then name the observation that would support it and the result that would make you reconsider. Avoid turning an interesting correlation into a sitewide rule.

You can read a repository without installing its code or granting it access to your accounts. Start with the research and its license. Run software only when you have a defined need, suitable permissions and a reviewed setup.

Repair a passage the buyer can use

A useful passage gives the answer, its conditions and enough evidence to check it. Read it outside the rest of the article. If “this,” “they” or “the platform” could refer to several things, name the subject. If a limit changes the answer, keep that limit in the same passage.

Consider a hypothetical CRM integration page. Its opening says: “Our AI integration connects every tool instantly and removes manual work.” The audit finds no support for “every,” “instantly” or “removes.” The page also fails to name the CRM, the trigger or the point where a person still checks the work.

For this example, suppose a documented test records three steps: a form submits a lead, the workflow checks required fields, and the CRM creates a contact. An incomplete record enters a review queue. A repaired passage could say:

The workflow sends completed enquiry forms to the CRM after checking the required fields. It creates a contact when those checks pass and sends incomplete records to a review queue. A person checks that queue before the team follows up.

This hypothetical workflow illustrates the edit. It makes no claim about a NuroSparx result. A real page should name its actual CRM, link to the relevant integration documentation and show the test conditions. Add a result only when the record supports it.

Then check the next paragraph. Does it explain duplicate handling, failure alerts or another decision the buyer needs? Replace a repeated benefit claim with that missing detail. A page becomes more useful when it answers the next practical question, even if the edit adds no new heading or statistic.

A reviewer circles the same architectural detail on a blueprint and on the page that cites it.

Keep structure tied to the content

Use headings, lists and tables when they help the reader complete the task. Put a direct answer under a question heading, then explain the conditions. Keep a comparison table only when you can fill its cells with meaningful, supported differences.

Google’s guidance for generative AI search calls for useful, original content that people can follow. A first-hand test, a precise explanation or a documented tradeoff gives the reader something to assess. Repeating a general checklist adds little evidence.

Audit structured data against the visible page. The author, product, service and FAQ details should describe what the reader can actually find. Correct an inaccurate field before expanding the markup. Treat the markup as a description of the page, and don’t assign it an invented citation score.

Review links the same way. A link to your search and content service belongs where the reader needs help carrying out the work. A related article belongs where it answers a follow-up question. Avoid adding a fixed number of links to each section just to satisfy a checklist.

Finally, check the page on a phone. Can a reader compare the table columns, read the labels in an image and follow a link without losing their place? Keep the main explanation in text even when a diagram also shows it. The image can clarify the process; the prose should still work without it.

Plan the measurement before the edit

Choose the outcome before you change the page. A crawler visit measures access. An answer that links to your page measures a citation in that sampled answer. An analytics session measures a visit. A saved, valid enquiry measures a business action. Record each under its own name.

For a citation check, save the exact question, engine, date, settings and answer. Count brand mentions separately from links to your domain. Keep the wording and test conditions steady when you repeat the check. If you change the question, begin a separate series so each comparison uses the same question.

For a page refresh, save the previous content and the preceding reporting period. Pick a relevant page you won’t edit as a comparison, and explain why it makes sense. It won’t create a controlled trial by itself, but it can reveal a wider traffic change that affects both pages.

A new page has no earlier performance at that URL. Mark its baseline as new. Compare its progress with a relevant group of new pages of similar age, when you have enough examples. Don’t describe growth from zero as proof that a particular writing technique worked.

We recommend checks at 7, 14 and 28 days after publication as an observation schedule. Use these dates to review the evidence. They do not promise that an engine will crawl, cite or rank the page by then. Record reporting delays and small samples before drawing a conclusion.

For traffic and enquiries, check the landing page and source alongside the form record. Exclude known tests. Our traffic-without-leads diagnosis covers the conversion checks that follow. More citations can justify further investigation; they don’t establish more revenue.

An analyst compares the source and search answer on two screens, then files an enquiry marked with the same shell.

Publish the smallest useful change

Close the audit with a specific edit and its evidence. “Improve authority” gives someone little to act on. “Replace the unsupported setup-time claim with the measured steps and link the test method” tells the writer exactly what to change and how to check it.

Save the current version, make the edit and inspect the rendered page before release. Check the title, canonical, links, images and form. After publication, confirm that the live page contains the reviewed text. Record the URL and publication date beside the observation dates.

If the page still leaves the buyer’s question unanswered, write a new supporting article with a distinct purpose. If the current page can answer it through one careful revision, start there. Keep the change small enough that you can describe what you learned when the results arrive.

Start with the page closest to a real buying decision. Our SEO, GEO and AEO audit can help you identify the evidence gaps and choose the first repair.

FAQ

What is a GEO content audit?

A GEO content audit reviews a page’s buyer question, access, evidence, structure and measurement plan. It produces specific changes that make the page more useful and easier to check. Keep crawl access, sampled AI citations, website visits and valid enquiries as separate outcomes so the report explains what each result proves.

Should I add schema before improving the article?

Correct inaccurate structured data when you find it, and make sure it matches the visible page. Review the article’s answer and supporting evidence alongside that work. A well-formed markup block cannot supply a missing test result, explain an unclear offer or establish a citation outcome on its own.

Can GitHub research tell me how to rank in AI answers?

GitHub research can give you methods and hypotheses to test. Read the sample, system, metric and limitations before applying a finding. A controlled benchmark, one site’s traffic study and a commercial assistant’s live answers measure different things. Preserve those distinctions when you decide what to change on your own site.

How often should I repeat the audit?

Repeat the affected checks after each meaningful edit. Use a planned observation schedule, such as 7, 14 and 28 days, to review results without changing the page every day. Revisit source facts when products or policies change. Those dates organize the work; they don’t guarantee when search systems will respond.

Should I publish a new page or update an existing one?

Update the existing page when it already serves the same reader and question. Consider a new page when the buyer needs a distinct answer that deserves its own URL. Check the current inventory first, define the new page’s purpose and link related pages without repeating their full explanations.

Free SEO, GEO and AEO Audit

See where you lose search and AI traffic

We check your rankings, AI answer visibility and technical issues, then send the fixes in priority order.

Your data is safe. We never share details without permission.

Free growth audit

Find out what's costing you leads

We check your site, ads and tracking and send the fixes that matter most within 24 hours. Your first 15 days are free on a new plan.