AI search measurement
AI Search Visibility Audit: A Practical SEO and GEO Checklist
Audit your AI search visibility with buyer questions, source checks and an SEO/GEO checklist. Includes worksheets to prioritize fixes and measure qualified leads.
On this page
An AI search visibility audit checks whether buyers can discover your business in AI answers, whether those answers describe it accurately, and whether a suitable buyer can take the next step. Start with real buyer questions, save the answers and cited sources, then compare them with verified facts about your offer. The useful output is a prioritized list of fixes with owners and evidence, rather than a single visibility score.
This guide is for founders and marketing teams deciding what to fix on an existing website. It covers the overlap between search engine optimization (SEO) and generative engine optimization (GEO), with worksheets you can copy into a spreadsheet.
What should an AI search audit measure?
Keep six observations separate. They describe different parts of the buyer's journey.
| Observation | Audit question | Evidence to retain |
|---|---|---|
| Mention | Does the answer name your business? | Complete answer and date |
| Citation | Does it link to your site or another source? | Exact cited URL |
| Accuracy | Are its statements about your offer correct? | Claim and approved fact source |
| Recommendation fit | Does your business meet this buyer's requirements? | Requirements and eligibility check |
| Referral traffic | Did someone visit from an AI experience? | Observed source and landing page |
| Business outcome | Did the visitor become a qualified lead or customer? | Defined conversion and qualification |
A mention can be inaccurate. A citation may link to a page without recommending the business. A recommendation can send someone to a broken form. Improving one observation does not establish improvement in the others.
For this checklist, a qualified recommendation means the business meets the hard requirements stated in the question. That is an audit definition, not a metric supplied by an AI platform. Decide your requirements before scoring answers.
1. Build a panel of real buyer questions
Choose one offer, audience and market for the first audit. A scheduling product for small teams and an enterprise scheduling platform may need different facts, comparisons and conversion paths.
Take questions from sales calls, enquiry forms, support tickets, site searches and customer interviews. A manageable starting panel might contain 12 questions; that is a working sample, not a statistically representative benchmark.
Include several kinds of question:
- Discovery: “Which scheduling tools suit a small service business?”
- Requirements: “Which tools include calendar sync and team permissions?”
- Comparison: “How does Example Scheduler compare with another option?”
- Branded verification: “Does Example Scheduler include SSO on its starter plan?”
- Negative fit: “Can Example Scheduler meet a requirement it does not support?”
These are fictional examples. Replace them with your customers' wording and real product constraints. Include location, budget or plan limits when buyers actually mention them; adding irrelevant constraints just makes the panel artificial.
Negative-fit questions matter. If an assistant recommends your product for an unsupported requirement, that is an accuracy problem even if the mention count rises.
2. Record answers in a repeatable way
Use the AI search experiences your buyers use. Record the product or search surface, visible mode, date, language, location context, exact question and whether web search was used. Start a fresh conversation for each independent question so earlier discussion does not silently supply missing facts.
Save the full answer, not just a screenshot of your name. Open the citations and retain the URLs. Mark an absent citation as absent; do not guess which page influenced an answer.
Repeat the same panel on a second occasion before treating one missing mention as a pattern. This is a consistency check, not a statistically significant experiment. Answers can vary between runs, users and product versions.
Use this answer log:
| Field | What to enter |
|---|---|
| Question ID and text | Fixed wording and buyer task |
| Run context | Date, surface, mode, language and location |
| Answer record | Full text or a retained screenshot |
| Mention and citation | Business named? Exact linked URLs? |
| Claim check | Correct, incorrect, incomplete or unverified |
| Fit check | Meets stated requirements, fails or unclear |
| Next step | Landing page and purchase/enquiry path |
If you calculate mention frequency, show the denominator: “named in 4 of 12 recorded answers on this surface.” Do not present that small sample as your share of every buyer conversation or combine different surfaces into an unexplained “AI rank.”
3. Check whether the right pages can be found
Before commissioning more content, check the pages that should answer the panel: your product or service page, pricing, documentation, comparisons, availability and contact page.
For each priority page, ask:
- Does the public URL load successfully without a login or security challenge?
- Is the page linked from a relevant, crawlable part of the site?
- Are indexing directives and the canonical URL consistent with the intended page?
- Are important eligibility facts present in readable page content?
- Do robots rules or CDN controls block the crawler you intend to allow?
Google says that supporting links in AI Overviews and AI Mode must be indexed and eligible to appear in Search with a snippet. Eligibility does not guarantee inclusion. Use Search Console to inspect priority URLs rather than assuming that a sitemap submission means a page is indexed.
For ChatGPT, OpenAI distinguishes OAI-SearchBot, used for search, from GPTBot, used for potential model training. Their controls are independent. A decision about training access should not be confused with a decision about search discovery. OpenAI also advises checking crawler access at the host or CDN, not only robots.txt.
Record an access issue separately from a content issue. A page containing the answer but blocked from the relevant crawler needs a different fix from an accessible page that omits the answer.
4. Compare every material claim with a fact source
Build a small reference sheet for facts that affect buying decisions. Assign an owner to each fact and identify its approved public URL. Internal knowledge helps verify the audit; public evidence helps a reader check the claim.
Useful fields include supported features, plan eligibility, service area, integrations, implementation requirements, availability and exclusions. Do not invent prices, customer results or product capabilities to fill a gap.
Here is a fictional finding:
| Item | Example audit entry |
|---|---|
| Buyer question | Does Example Scheduler include SSO on Starter? |
| Recorded answer | Says SSO is available on Starter |
| Approved fact | SSO is available only on Enterprise |
| Evidence gap | Feature overview says “SSO available” without a plan limit |
| Proposed fix | Add plan eligibility beside the feature claim |
| Owner and check | Product owner verifies; marketing updates the page |
The evidence gap is a hypothesis about what to fix, not proof that the feature page caused the answer. Inspect the actual cited pages where present, plus other relevant public pages, before deciding.
Publish an explicit statement such as “SSO is available on Enterprise; it is not included in Starter” when that is the verified fact. Correct stale versions on pages you control. If an external directory is wrong, request a correction and track whether it was accepted.
5. Improve pages around unanswered buying decisions
Map each unresolved question to the most useful existing page. Add a new article only when the question deserves a separate explanation. Publishing several pages that repeat the same answer can leave buyers unsure which source is current.
A useful answer normally states:
- What the product or service does, using the buyer's language.
- Who qualifies, including important exclusions.
- What evidence supports the claim and where it can be checked.
- What the buyer should do next.
Keep important qualifications next to the claim they limit. A comparison table should distinguish a verified absence from “not confirmed.” A service page should name the locations actually served. A case study should explain its method and permission to publish, rather than borrowing an unsupported outcome.
This is the shared work of SEO and GEO: making useful information accessible and understandable. Google's AI optimization guidance emphasizes normal SEO foundations and distinctive, useful content. It does not prescribe special text chunking or an ideal word count, and says it does not use llms.txt for ranking. Treat the structure here as a tool for readers, not a citation formula.
6. Prioritize fixes by buyer impact
Start with errors that could disqualify or mislead a buyer, then access problems and missing decision evidence. An extra mention is less useful if the answer still sends the wrong customer to your sales team.
Copy this change log into your audit sheet:
| Finding | Priority | Owner | Completion evidence |
|---|---|---|---|
| Incorrect plan eligibility | High | Product + content | Verified limit on live page |
| Priority page blocked | High | Developer | Access check and URL inspection |
| Unsupported comparison claim | High | Subject owner | Claim evidenced or removed |
| Missing implementation detail | Medium | Delivery + content | Useful answer on relevant page |
| Broken enquiry form | High | Website owner | Successful test submission |
Adjust the priorities to your business. Keep the original answer, proposed correction, affected URL and publication date together. That gives the next audit a record of what changed.
Before closing an issue, follow the buyer's path on mobile: open the cited or intended page, check the offer, and complete the relevant purchase or enquiry flow using an appropriate test process. Verify the confirmation and routing, not just that the button can be clicked.
7. Recheck answers and measure business outcomes separately
After publishing corrections, rerun the original questions with the same recorded context. Allow for discovery and indexing delays. Keep an unchanged question panel for comparison; log new questions separately.
Compare factual accuracy, qualified recommendations and citations against the saved baseline. If an answer improves, record the observation without claiming that a particular edit caused it. Other sources and platform behavior may have changed too.
Use the reporting available for each surface. Google documents a Generative AI Search performance report for AI Overview and AI Mode impressions, with its own scope and limitations. Treat platform reports as evidence about that platform, not as a complete record of AI discovery everywhere.
In website analytics, inspect the AI referral sources and campaign parameters you actually observe. Track landing pages, conversions and lead qualification. Some journeys will not carry usable attribution: a buyer may read an answer and return directly later. An optional “How did you hear about us?” field can add context, but is not precise causal measurement.
Report three things side by side: what changed on the site, what the sampled answers did, and what buyers did. Keep uncertainty visible when the evidence does not connect them.
Common questions about SEO and GEO audits
Is an AI search audit different from an SEO audit?
They overlap on access, indexing, internal links and useful content. An AI search audit also checks the statements and recommendations inside sampled answers against the buyer's requirements. A technically healthy page does not establish that an assistant describes the business correctly.
Can I audit AI visibility without a paid tool?
Yes. For a focused first audit, use a spreadsheet, manually retained answers, public source pages and your existing search and analytics tools. Paid tools may make repeated collection easier; inspect their question coverage and definitions before relying on a headline score.
Does completing this checklist guarantee AI citations?
No. The checklist helps identify access barriers, missing evidence, inaccurate claims and broken buyer journeys. Search systems still choose which sources to show. Indexing, citation, recommendation and sales are separate outcomes, and none is guaranteed by completing an audit.
What should an agency deliver after the audit?
A useful handover includes the question panel, retained answers, verified fact sheet, prioritized findings, named owners and a recheck plan. If you need implementation support, use our GEO agency selection guide to assess the scope and evidence to request, or see how Findfront works.