AI visibility describes whether and how a business, product or website appears in AI-generated answers you observe. Being named, being cited and being recommended are distinct outcomes. An answer can do all three and still get a business fact wrong.
Start by asking what you want to learn. If you want to know whether an assistant suggests you for a buying task, inspect the role your business plays in the answer. If you want to know whether the answer gets your returns policy right, check the claim against the policy itself. Both observations are useful, and they answer different questions.
Four signals you can read in a single answer
Use this rubric to interpret one answer. Record each signal separately and keep the supporting sentence, URL or business source with the observation.
| Signal | Evidence to record | How to interpret it |
|---|---|---|
| Mention | The brand name or a verified alias in the answer body. | The business was named. Record a source-card appearance separately. |
| Citation | The displayed source link or reference. Note whether the URL belongs to your business or another publisher. | The platform displayed a source. Open it to check whether it supports the specific statement. |
| Recommendation | Wording that puts the business forward for the task. Preserve conditions such as “if it serves your location.” | The answer presented the business as a possible fit. A background example or source credit alone has a different role. |
| Accuracy | The exact assertion and the approved policy, specification or record you checked it against. | Mark each claim correct, incorrect or unverified. A link can accompany an incorrect claim. |
This is a reader-facing rubric. If a tool reports a composite visibility score, check which signals it includes and what it counts before comparing that score with another report. FoundUp’s public methodology distinguishes brand mentions, links to your site and accuracy of checked claims.
Illustrative example: a recommendation with an inaccurate detail
Imagine a fictional office furniture supplier, Harbor Desk. An AI answer says, “Harbor Desk is an option if you need installation outside business hours. Its catalogue lists the available desk sizes.” The answer displays a link to the supplier’s own catalogue. The current installation policy says its crews install during business hours only.
Record a mention, an owned citation, and a conditional recommendation. Mark the installation claim incorrect. The catalogue link may help a buyer check desk sizes. The relevant evidence for installation hours is the current installation policy.
The difference matters to the buyer. A team measuring only recommendations might count this answer as a success. Someone who needs evening installation may still waste time sending an inquiry. The useful action is to check where the wrong detail could be coming from, then correct any inaccurate public information the business controls. The example is illustrative and represents neither a live response nor a customer result.
When you cannot see the full evidence
If the answer is complete and your website only appears in a source card, record the citation and mark the body mention absent. If the platform displays no accessible source panel and you cannot tell whether sources were provided, mark citation status unknown. If the platform returns an error or an incomplete response, record the run separately rather than counting it as an absence. For a shortlist where some names are just references, read the wording around each name before counting recommendations. Reading an AI procurement shortlist
Why visibility can look different from one answer to the next
A question about “suppliers that install after hours” has a different purpose from “What is Harbor Desk’s installation policy?” Changing the wording, language, market context or interface can change the answer you see. Repeating the same question can also produce different results. That makes one response a useful lead for inspection and a weak basis for a trend. Schulte and colleagues’ repeat-observation study
Your website, directories, partner pages and independent sources may describe different versions of the same offer. If a consequential claim is wrong, check the official page and any visible sources. Note what you can verify and what remains uncertain. An absent mention alone does not reveal which information the system used or why it omitted a business.
Establish a baseline you can interpret
Start with a few questions tied to the same audience and decision. Keep discovery questions separate from branded fact checks. The former can show whether an answer puts your business forward without being given its name. The latter are better for checking whether specific information is right.
Write down the exact wording, platform, date, language and intended market context. Save the answer and available source URLs. If you cannot confirm a setting such as location, record that uncertainty. Keep question selection and conditions comparable when you revisit the result. You can do this with a spreadsheet. The important part is retaining enough context to explain what the observation means.
AI Visibility Baseline Template
Copy the header into a spreadsheet and add one row per answer. The sample row is Illustrative. It uses the fictional Harbor Desk answer above, with a reserved example domain and no live platform observation.
Prompt,Platform,Market,Language,Date,Mention,Citation,Recommendation,Accuracy,Citation URL,Claim Checked,Evidence Source,Sampling Status
"Which office furniture suppliers offer after-hours installation?",Illustrative assistant,US (intended),English,2026-11-12,Yes,Owned,Conditional,Incorrect,https://harbordesk.example/catalogue,Harbor Desk installs outside business hours,Fictional Harbor Desk installation policy,Illustrative - valid
Record whether the citation is owned or third-party, and preserve the displayed URL separately. Use “none” only when the complete source display shows no citation. Use “unknown” when you cannot inspect it. For a conditional recommendation, keep the relevant wording with the saved answer. Use Claim Checked for the specific assertion and Evidence Source for the policy or record used to verify it. A citation can point to a different source. Mark the Sampling Status valid, incomplete or failed. Leave signals unscored for failed runs.
Read the role before you count a recommendation
Compare “Harbor Desk makes reception desks” with “Consider Harbor Desk if you need a desk made to your floor plan.” The first names the company. The second puts it forward for a specific task. If you want to compare recommendations over time, compare answers to questions that actually ask for options. Fact-check questions belong in their own group.
Decide what to fix first
Use the evidence you have to choose one check your team can complete.
- A wrong fact could waste a buyer’s time. Verify the current fact and check the page where a buyer would expect to find it. If a public description is incomplete, correct that detail and retain the old wording for comparison.
- The right source is shown but the recommendation is qualified. Inspect the qualification. If the answer says buyers should confirm installation hours, give readers a direct, accurate policy on the appropriate page.
- The brand is missing from one discovery answer. Record the observation and check comparable questions before commissioning changes. If the same buyer need keeps appearing and you have no page that serves it, investigate that gap.
Review changes across three separate records
Track answer observations against the same questions and conditions. Keep search performance such as impressions and clicks in its own record. Review business results such as suitable inquiries or orders separately, using attribution only when you can support it. A more accurate answer can be useful even when your brand is mentioned less often, and a rise in citations does not establish that sales changed.
Start with one question that matters and one claim your team can verify. If manual checks become difficult to repeat, review how FoundUp scopes selected questions, reports evidence and handles follow-up. AI visibility monitoring scope