If you’re using an AI model like Claude or ChatGPT to audit your content for SEO, GEO, or AEO, it's important to verify that its recommendations are accurate, up to date, and relevant to your website before assigning work to your team.
In new research, we found that several AI models might recommend adding code for a Google Search feature that was eliminated months ago. They might also recommend a technical change without evidence that it improves rankings or AI citations.
In May 2026, Google stopped sharing FAQ-rich results, the expandable questions that appeared beneath some search listings. In our new analysis, four popular AI models included FAQ content or FAQ-structured data among their top five recommendations for a men's sneaker brand.
This post helps identify these and other potential problems before your developers spend hours implementing fixes that won’t increase your search results.
Follow these six steps to help fact-check your AI SEO audit and determine which recommendations deserve your time.
Step 1: Give the model a specific page to examine and goal
Don't use a prompt like "Audit my website for SEO."
A model needs to know the page you want to improve (and why), what consumers are likely searching for, and what you want the page to accomplish, like a purchase, a consultation, or an email sign-up.
For example, if you manage an online sporting goods store, you might want to improve a category page's visibility for searches like "best men's sneakers for new runners."
Begin by giving the model the page URL and the target search query. Use a model with web access enabled or provide the page content and crawl data if the model can’t inspect the website.
You can use this prompt as a template:
Review [PAGE URL] for the search query "[TARGET QUERY]."
Identify the five most important SEO improvements.
For each recommendation:
Identify the exact problem and show the page evidence.
Explain what the proposed fix is expected to improve.
Link to the current Google Search documentation supporting the recommendation.
State whether the fix addresses a confirmed problem or is a hypothesis that needs testing.
Rank the fixes by expected impact and explain your ranking. If you cannot inspect the page or verify a claim, say so.
The model should provide you with recommendations tied to a problem identified on your page.
Here’s an example of a bad recommendation: "Improve your structured data to increase AI visibility."
Here’s an example of a good recommendation: "The page has no valid breadcrumb markup. Check if adding BreadcrumbList structured data would make the page eligible for Google's breadcrumb search appearance."
The second recommendation identifies a likely problem, an outcome, and a verifiable Google feature.
Step 2: Make the model explain what each fix is supposed to accomplish
Before you implement an AI recommendation, consider this question:
What result should we expect if we make this change?
Consider this recommendation to add FAQ schema involving these elements:
FAQ content: Questions and answers that consumers can read on your website.
FAQ schema: Code that identifies those questions and answers for systems that process structured data.
FAQ-rich results: A Google Search feature that displayed expandable questions beneath some search listings.
Adding content that answers specific questions, like "Are these shoes suitable for flat feet?”, helps users choose the right product for their needs. However, implementing FAQ schema won’t restore retired FAQ-rich search results. In its guide to generative AI features, Google notes that AI Overviews and AI Mode don’t require specialized schema markup. Understanding this difference is important because improving page content and adding technical markup serve different objectives.
You can copy and paste this follow-up prompt:
“For each recommendation, provide the expected benefit to our brand, such as:
Improved information for our customers
Increase the percentage of indexed pages
Increase eligibility for specific Google Search features
Increase organic search traffic
More citations within AI-generated responses”
Clarify which benefits are backed by Google's documentation and which are unsupported. Avoid allowing the model to merge multiple benefits into a single, unverified claim.
Step 3: Check the recommendation against Google's current documentation
Sometimes, a model can answer a question correctly while suggesting an outdated or low-value tactic.
Our analysis uses common SEO/GEO questions from ChatGPT, Claude, Gemini, and Perplexity. Each model responded to four prompts, including a request for its top five SEO fixes for a men's running shoes category page.
All four included FAQ code or FAQ content in their top five recommendations.
Three of the four models described Google's earlier restrictions on FAQ-rich results without mentioning its removal. Perplexity identified the removal date but included FAQ-structured data in its recommendations.
In several instances, the models knew a feature was removed and recommended work without establishing a benefit to its user.
Here are a few ways that you can check a model's recommendation:
Search the page for the feature the AI recommends, such as "FAQ-rich results."
Check the most recent change and its effective date.
Read the current documentation to confirm whether the feature applies to your brand.
Compare Google's guidance with the benefit the model stated
As a best practice, don’t rely on a model's confidence, the number of models in agreement on a topic, or the presence of a citation, and always verify the model’s sources.
Step 4: Put verified fixes ahead of unsupported recommendations
After your verification, return to the audit and reorder the model's recommendations.
Assume your page has three common SEO issues:
Google cannot index the page because of an unintended noindex directive.
The title tag doesn't describe the page's products.
The page doesn't have FAQ structured data.
Start by fixing any index issues because if a page isn’t indexed, it’s invisible in Google’s search results.
Next, improve the title tag so it describes your category and matches the content on the page.
Use FAQ code only if you can identify a supported reason for adding it.
You can use the following prompt (or a variation of it) to review the order of work:
“Re-rank the recommendations using only verified issues and supported benefits.
Prioritize problems that prevent crawling or indexing, followed by problems that limit the page's ability to answer its target search query.
For each fix, explain why it should come before the next one. Identify recommendations that lack enough evidence to justify implementation.
Your final instructions should tell your team what to fix and in what order.
Step 5: Examine the model's recommendations when the benefit to your brand is uncertain
Sometimes, Google supports a technical feature that a model might recommend without providing evidence that it will improve your specific results.
Structured data is a good example.
An Ahrefs study published May 11, 2026 examined 1,885 pages that added JSON-LD structured data and compared them with 4,000 similar pages that didn't.
Relative to the comparison pages, AI citations changed by:
Google AI Overviews: Down 4.6%.
Google AI Mode: Up 2.4%.
ChatGPT: Up 2.2%.
The changes in AI Mode and ChatGPT weren’t statistically significant. The AI Overviews decline was statistically significant, but researchers couldn't establish schema as a cause.
The study combined several types of schema, so the results don't isolate FAQ markup. The takeaway is that a feature can be valid without increasing AI citations.
If you want to find out what information might improve your website's AI visibility, try this simple analysis:
Set up:
Choose five to ten pages with AI citations.
Choose five to ten similar pages with comparable citation activity.
Document how often each group appears as a cited source in AI answers during a set period of time.
Make the proposed change to the first group; don’t modify the comparison pages.
Avoid making major changes to either group during your evaluation.
Measure citations again over the next 30 days and compare the differences between both groups.
For example, suppose the updated pages gain 20% more citations while the unchanged pages gain 18%.
The updated pages outperformed by two points. That difference isn't enough to show that the new changes were effective.
If the updated pages gain 25% while the unchanged pages lose 10%, you have cause to investigate the change's possible effect.
An analysis involving 10 or 20 pages won't prove causation because each model’s results vary.
However, comparing modified pages against an unchanged control group provides more insight than tracking just the modified pages.
Step 6: Keep track of outdated advice
After your audit, save its recommendations and the information you used to decide what changes to make.
Here’s an example of results worth documenting:
| Field | Example |
|---|---|
| Recommendation | Add FAQPage schema |
| Intended benefit | Google FAQ-rich results |
| Official source | Google Search Central |
| Source last checked | November 16, 2026 |
| Finding | Google removed FAQ-rich results |
| Decision | Don't implement for this purpose. |
| Next review | December 2026 |
Repeat the analysis when you run a new audit, particularly for recommendations involving Google features like structured data and AI search.
How to Know When Your AI SEO Audit Is Accurate Enough To Act On
An AI SEO audit is ready to act on when its recommendations can answer these questions:
What information is keeping the page out of your desired search results? What information supports the model's recommendation for improvements? How will we know if the model's recommendations will work?
If the audit can't answer those questions, don't assign new work until they can. Use AI to identify possible problems and speed up your research, but remember to verify its recommendations and prioritize issues before committing valuable developer time.
At 5WPR, we help brands evaluate how AI search systems discover, describe, and share their content, then develop SEO and generative engine optimization programs based on evidence rather than unsupported recommendations. We help our clients build brand authority across the platforms where decisions happen: ChatGPT, Claude, Perplexity, Gemini, and Google AI Overviews, alongside earned media, digital, and influencer channels. 5W combines public relations, digital marketing, and proprietary research to help clients measure and grow their presence in AI-driven discovery.






