Pick a set of questions a buyer might ask, save the answers you see, and run the same questions again after you’ve checked a finding or changed a page. With a spreadsheet and access to an AI answer interface, you can build a baseline that tells you where your brand appeared, how it was presented, and which claims are worth checking.

This guide uses Cedar Finch Workshop, a fictional reception desk maker. Its prompts, answers, sources, dates, and results are synthetic. Use the example to see how the method works, then substitute your own questions and evidence.

Quick start

Have ready: a spreadsheet, access to one AI answer interface, a few buyer questions, and a reliable reference for one fact about your business, such as a shipping policy.

  1. Run the same set of questions in one interface and save what it says and links to.
  2. Mark where your brand is named, offered as an option, or cited. Check any specific business claims you can verify.
  3. Choose one finding to investigate, record your decision, and repeat the questions on a date you set.

You’ll finish with a baseline you can check, a record of any action you took, and a second set of answers you can compare. Start with a panel small enough to review by hand.

1. Build a prompt panel around buyer decisions

Start with a buyer and a decision. A clinic manager looking for a custom desk might first ask which makers serve small clinics, then ask whether a supplier can work from an existing floor plan. Those are different decisions, even though both questions concern reception desks.

Give each prompt an ID and label its purpose. Use unbranded discovery prompts to see which options an answer offers without naming yours. Use branded fact checks to see whether the answer describes your business accurately. In your own panel, record whether a question came from a customer, your research, or a tool. Preserve the exact wording you actually run.

Synthetic prompt panel — Cedar Finch Workshop

IDExact promptTaskWhy it stays
P1Which US makers offer custom oak reception desks for small clinics?Unbranded discoveryTests whether Cedar Finch appears in a shortlist for a specific buyer.
P2Which US suppliers can make a reception desk from my floor plan?Unbranded selectionTests a requirement the clinic shortlist doesn’t address.
P3Does Cedar Finch Workshop ship reception desks to Alaska?Branded fact checkTests a delivery claim against a policy.
P4Can Cedar Finch Workshop build a desk from a customer floor plan?Branded fact checkTests what the answer says about a specific supplier’s intake process.

All four questions were written for the example. Leave out “best custom clinic reception desk makers” if it repeats P1’s decision. Leave out “what is AI visibility?” because it serves a different reader. P2 and P4 both mention floor plans, but only P2 asks the answer to find a supplier.

Freeze the wording, then choose the interface, market, language, and repeat schedule. Start with a panel you can rerun and inspect. If you later add or reword questions, record the change and retain the original IDs for comparisons. A spreadsheet is enough to start. Download the prompt tracker.

2. Run the baseline and save the answer evidence

Ask one question and save the evidence before moving on. Paste the full answer into a dated note or text file. If a screenshot helps preserve context, include the prompt, answer and visible source panel. Copy the actual source URLs into the sheet when available. Use the same ID in the sheet and the saved file, such as B-P1-a1.txt, so you can find the evidence later. If you can save only an excerpt or screenshot, note what it captures and whether the rest of the answer or the sources are missing.

Record the prompt ID, interface or API, model if shown, timestamp with time zone, market and how you established it, language, search mode, and session conditions. For settings you cannot observe, write unknown. For settings you can observe but cannot change, record the observed value and mark it uncontrolled. If you intended to sample the US but the interface does not confirm the region, record US as the target and the actual region as unknown. Keep browser and API results in separate groups.

Each time you submit a question is an attempt. A slot is the answer you planned to collect for one prompt on one surface in one round. If an attempt fails, a retry is another attempt for that same slot. Set the retry rule before collecting answers. Here, a failed execution gets one retry within five minutes under the same conditions. The first valid answer fills the slot. A readable answer with an unavailable sources panel still fills the slot, with citation status marked unknown.

Synthetic baseline highlights — Cedar Finch Workshop

  • P2: The first submission timed out. The retry produced a usable answer. Keep both attempts and score the retry once.
  • P3: The answer was readable and the sources panel was unavailable. Keep the answer and mark its citation status unknown.
  • P4: Both attempts timed out. Count both failures and leave the planned slot unfilled.

That gives three valid answers from four planned slots and three failed attempts out of six. See the full synthetic attempt log for the timestamps, answers and source records.

3. Code mentions, recommendations, citations, and accuracy separately

Read the answer body first. Then check the source links and compare any business facts with a dated reference. Record the excerpt and URL that support each decision so another reviewer can follow it.

  • Mention: Look for the brand in the answer body. Confirm an alias before counting it.
  • Recommendation: Check whether the answer offers the brand as an explicit fit or a conditional option. A background reference and a factual statement have different roles.
  • Owned citation: Check whether a visible source link goes to a verified page the brand owns.
  • Accuracy: Compare checkable brand claims with a dated fact record. Record an unresolved claim as unknown and an answer with no eligible claim as not applicable.

The full coding rubric gives the rules for yes, no, unknown and not applicable.

In the synthetic baseline, P1 presents Cedar Finch as a fit and cites its clinic desks page. You can check that the company offers custom oak desks for clinics. Whether it is a good fit is a recommendation judgment. P2 presents Cedar Finch as an option if it confirms the dimensions. The answer gives it a conditional recommendation and cites a third-party URL. It leaves the actual floor-plan policy unconfirmed, with no factual claim to score.

P3 names Cedar Finch and gives a factual answer about Alaska. The policy limits delivery to the contiguous 48 states, so the answer is inaccurate. The missing panel leaves the owned citation unknown. For a phrase such as “Cedar Finch might be relevant,” look for whether the answer actually presents the company as a candidate. Leave the role unknown when the context is ambiguous. See how candidate roles affect the count.

Separate synthetic example — brand absence. This example sits outside Cedar Finch’s ten attempts. Suppose a complete answer to a supplier question reads “Try North Pine Fixtures.” Here is how to record Cedar Finch:

What you can verifyWhat to record
The full answer and source panel show no Cedar Finch name or owned URL.Mention No. Recommendation No. Owned citation No.
The body is unchanged and a source card links to https://cedarfinch.example/spec-guide.Mention No. Recommendation No. Owned citation Yes. Verify ownership of the linked page.
The body is complete and the sources panel is unavailable.Mention No. Recommendation No. Owned citation Unknown.
The submission times out and no answer is available.Record a failed attempt. Follow the retry rule and leave all answer outcomes unscored unless a valid retry fills the slot.

All three readable variants make no eligible claim about Cedar Finch, so accuracy is not applicable. The separate case leaves the original coding table intact.

4. Calculate the baseline and compare like with like

Report coverage first: valid slots divided by planned slots. Report failed attempts separately: failed attempts divided by all attempts. Then calculate each scored outcome as yes ÷ (yes + no) among valid slots. Report unknown and NA separately so readers can see why denominators differ.

Use the recommendation coding in three ways:

  1. Role in one answer. P2’s baseline answer presents a conditional option. The retest mentions Cedar Finch for background, so its recommendation code is no.
  2. Recommendation occurrence across the full panel. Count valid slots with a recommendation and report coverage of the role coding. In this example every valid slot has a scorable role, so the rate is 2/3 at baseline and 1/4 at retest. If any role is unknown, keep it visible and report the observed count out of all valid slots alongside the scorable rate. The full-panel proportion describes this mix of prompts, including fact checks.
  3. Discovery and selection Recommendation Rate. Use only prompts selected in advance to find or shortlist providers. Count explicit and conditional recommendations in the numerator and scorable discovery and selection slots in the denominator. Cedar Finch’s P1 and P2 yield 2/2 at baseline and 1/2 at retest. P3 and P4 ask for facts about a named business, so they belong in the fact-check group.

Synthetic round totals — Cedar Finch Workshop

MeasureBaselineRetest
Valid slots / planned3/44/4
Failed attempts / attempts3/60/4
Mention3/34/4
Full-panel recommendation occurrence2/31/4
Discovery and selection recommendation rate2/21/2
Owned citation1/2, 1 unknown3/4
Accuracy1/2, 1 NA3/3, 1 NA

For change over time, match prompt IDs and conditions. Include a pair for a metric only if both answers are scorable for it. In the synthetic example, P4 has a valid answer only on the retest. It adds coverage and a scorable fact to that round, so leave it out of the paired comparison.

Synthetic paired comparison — Cedar Finch Workshop

OutcomeEligible prompt pairsBaseline → retest
MentionP1, P2, P33/3 → 3/3
Full-panel recommendation occurrenceP1, P2, P32/3 → 1/3
Discovery / selection recommendation rateP1, P22/2 → 1/2
Owned citationP1, P21/2 → 2/2
AccuracyP1, P31/2 → 2/2

P3 has a citation at retest, but its missing baseline sources panel leaves no paired citation result. P2’s retest citation counts in the citation comparison even though the answer stopped recommending the brand. The two measures answer different questions.

AI answers can vary across runs, prompts, and time. A 2026 preprint by Schulte and colleagues supports repeating observations. The retry window, scoring rules, and matched-panel procedure here are choices for this workflow. Treat the counts as descriptions of the collected sample. Don’t Measure Once: Measuring Visibility in AI Search.

5. Decide what to check, then rerun the panel

Use a signal to choose an evidence check. Decide whether to change a page only after you can name the specific issue you verified.

SignalEvidence checkPossible action
Brand absent from discovery answersConfirm that the body was complete. Check other comparable answers and whether the offer actually fits the question.If an eligible offer is poorly described on its page, clarify it. If the sample is sparse, collect another comparable round.
Fewer owned citationsCheck source-panel completeness, matched prompts, exact URLs and whether a URL moved.Correct a confirmed broken reference you control, or repeat collection if source coverage changed.
Answer gives a wrong business factCompare the exact claim with the dated policy and current page copy.Clarify a verified ambiguity and log the change for retest.
Answer changes between roundsCompare prompt wording, interface, settings and site changes. Review the saved text side by side.Document the difference and repeat the test before attributing it to any one change.

Start with a claim you can verify. In the synthetic Cedar Finch example, B/P3/a1 says the company ships to Alaska. Check the dated shipping policy and the public page before deciding what to change. The policy excludes Alaska, while the page says “US delivery available.” That is a concrete ambiguity the site owner can fix. The missing source panel leaves the origin of the wrong answer unknown.

Synthetic change and retest record — Cedar Finch Workshop

  1. November 2, B/P3/a1. Compare the Alaska claim with policy POL-03, which limits shipping to the contiguous 48 states. Check the broad wording on https://cedarfinch.example/shipping.
  2. November 4, CH-01. The fictional operations owner approves replacing “US delivery available” with “We ship to the contiguous 48 states. Alaska delivery is unavailable.”
  3. November 9, R/P3/a1. Under the same recorded protocol, the answer gives the correct policy and lists the shipping page as a source.

Rerun the same prompts on the same surface and record any changes to the interface, model, settings, or site that you know about. In this synthetic retest, P1 repeats the explicit recommendation and its owned citation. P2 says, “An independent shop could work. For background, see Cedar Finch Workshop,” and lists the owned https://cedarfinch.example/spec-guide URL. P3 says, “No. Cedar Finch Workshop ships only to the contiguous 48 states,” and lists the owned shipping URL. P4 confirms that Cedar Finch accepts floor plans subject to dimensional review. Its complete source panel lists no sources. All four retest attempts were valid on the first try.

The P3 answer and source changed after the simulated edit. Record that sequence and keep observing before drawing a conclusion about what caused the change. For your own site, start with one prompt whose answer points to a fact your team can check. Save the evidence, record the decision, and schedule the repeat run. Keep monitoring by hand while the panel is small, a named teammate can rerun it on schedule, and someone can inspect the saved answers and follow up. Consider managed monitoring when multiple surfaces or markets make collection hard to sustain, runs are repeatedly missed, or findings need regular review across teams. If that is where your team needs help, explore AI visibility monitoring.

Appendix: Synthetic case records

The log and coding tables below contain fictional teaching data. The prompt IDs join to the exact wording in step 1. Both waves use fictional browser surface A, simulated US region selected, en-US, web search on, and a fresh signed-out session per attempt. Model version is unavailable. All dates are in 2026 and all times are UTC.

Full synthetic attempt log

Slot / attemptTimeStatusSaved answer or failure detailSources capture
B/P1/a1Nov 2 14:00ValidCedar Finch Workshop is a good fit for clinics seeking a custom oak reception desk.Owned: https://cedarfinch.example/clinic-desks
B/P2/a1Nov 2 14:01TimeoutNo answer.Not applicable
B/P2/a2Nov 2 14:03Valid retry of a1Consider Cedar Finch Workshop if it confirms your dimensions. Another option is an independent shop.Third party only: https://tradejournal.example/shop-directory
B/P3/a1Nov 2 14:04ValidYes, Cedar Finch Workshop ships desks to Alaska.Panel unavailable. Citation status unknown.
B/P4/a1Nov 2 14:05TimeoutNo answer.Not applicable
B/P4/a2Nov 2 14:07Timeout, retry of a1No answer.Not applicable
R/P1/a1Nov 9 14:00ValidCedar Finch Workshop is a good fit for clinics seeking a custom oak reception desk.Owned: https://cedarfinch.example/clinic-desks
R/P2/a1Nov 9 14:01ValidAn independent shop could work. For background, see Cedar Finch Workshop.Owned: https://cedarfinch.example/spec-guide
R/P3/a1Nov 9 14:02ValidNo. Cedar Finch Workshop ships only to the contiguous 48 states.Owned: https://cedarfinch.example/shipping
R/P4/a1Nov 9 14:03ValidYes. Cedar Finch Workshop accepts floor plans, subject to dimensional review.Complete sources view, no source listed.

Full coding rubric

OutcomeYesNoUnknown or not applicable
MentionThe target is named in the answer body. Count an alias if you can confirm the identity.The complete body omits the target. A name in a source card alone is outside this measure.Identity or answer completeness is unresolved.
RecommendationThe answer presents the target as an explicit fit or a conditional option. Record which role it played.The answer mentions the target as background, states a fact about it, or leaves it out.The role or identity is unclear.
Owned citationA visible source links to a verified URL owned by the target. Keep the original link.The complete source view shows no owned URL.The source view is incomplete or ownership cannot be confirmed.
AccuracyThe answer makes at least one eligible brand claim, and all eligible claims agree with the dated facts you checked.An eligible claim contradicts the verified facts.Mark unknown if a material claim is unresolved and no contradiction is verified. Mark NA if there’s no eligible claim.

Original slot coding

Synthetic answer coding — Cedar Finch Workshop. Y = yes, N = no, U = unknown, NA = not applicable.

SlotMentionRecommendation roleOwned citationAccuracy
B/P1YY, explicit fitYY
B/P2YY, conditional optionNNA
B/P3YN, factual referenceUN
B/P4NANA, failedNANA
R/P1YY, explicit fitYY
R/P2YN, backgroundYNA
R/P3YN, factual referenceYY
R/P4YN, factual referenceNY

What Is AI Visibility?