Follow a correction through one audit record
Write down what changed in the source and what changed in later answers. Keep those two results separate.
Worked audit · fictional example
One package. Two prices. Trace the source.Cited business profile · outdated
$140
- Package
- Apartment cleaning
- Coverage claim
- Seattle-wide
Needs correction
Operations record · approved
$160 total
- Package
- Two-hour standard clean
- Supported area
- ZIP 98101
Use these facts
Correct the profile. Check the service page and booking total against the approved package, then repeat the original question.
Compare the assistant's answer with the approved cleaning package, identify the stale price or coverage claim, and correct the source before repeating the same question.
After correcting the profile, repeat the question. Record factual accuracy and recommendations separately; the answer may not refresh immediately.This filled record illustrates the worksheet using a fictional cleaning business. The relative paths identify example evidence; they are not live customer pages or a PingPlus campaign result.
| Worksheet field | Completed example |
|---|---|
| Question and context | Who offers a two-hour apartment clean in ZIP 98101 on Saturday for under $180 total? Fresh conversation, English, US market; log the actual assistant and mode used. |
| Observation evidence | Save the dated answer and its cited URLs. In this teaching scenario the answer mentions and recommends the business, citing its older business profile rather than its website. |
| Cited source and incorrect fact | Business profile, saved as evidence/profile-before: $140 and Seattle-wide coverage. |
| Approved source and fact | Current /services/apartment-cleaning page plus operations' dated price and coverage record: $160 total for the stated package in ZIP 98101; Saturday slot subject to booking confirmation. |
| Priority and correction | High: misleading price and coverage. Correct the profile's package price and supported ZIPs; verify the service page and booking check agree. |
| Owner and acceptance check | Operations owns approved facts; marketing owns the profile correction. Save the updated source and test both a supported ZIP and an unsupported one. |
| Repeat-check result | Not yet run. Repeat the original question and separately record source accuracy, citations, and recommendations. |
- Capture: keep the answer, cited URL, requested ZIP, and conflicting price together. Operations confirms the approved $160 total and coverage.
- Correct: update the profile, align the service page, and link to the coverage check. Keep dated evidence of the correction and its owner.
- Recheck: repeat the same question and record whether the price is accurate, whether the source changed, and whether the business is recommended.
If the next answer gives $160 but omits the business from its shortlist, accuracy improved; visibility did not improve in that observation. If it still cites the old information, check whether the source actually updated and whether the answer exposes a different URL. Do not repeatedly rewrite the accurate page just because a platform has not refreshed its answer.
Review after a meaningful source change and on a regular schedule your team can maintain. Track unresolved issues, completed corrections, factual accuracy, and comparable appearance counts. Use traffic and qualified customer outcomes as separate measures.
Choose questions that reveal a buying decision
Build a small question set from actual customer enquiries. Include the constraints that would change whether your business is suitable.
A useful audit begins before you open an assistant. Choose one offer and list the questions customers ask before buying, booking, or requesting a demo. Keep the source of each question: a sales call, a support message, a search query, or an editorial scenario. That prevents your own guesses from quietly becoming customer research.
For a fictional cleaning service, a useful question is “Who offers a two-hour apartment clean in ZIP 98101 on Saturday for under $180 total?” A broad “best cleaner” question leaves the team guessing which part of the offer mattered. Add a brand question, a comparison question, and a deliberately unsuitable request, such as an area the business does not serve.
Start with five to ten questions that cover different decisions. This is a manageable editorial starting point, not a statistically representative sample. Preserve the wording for repeat checks. If you later change the question, start a new series rather than treating it as the same observation.
Use the guide to choosing AI search queries to decide which requests deserve attention. Audit fit before chasing a mention.
Record what happened before interpreting why
Save the question, context, answer, sources, and observation date. A mention, a citation, and a recommendation are different observations.
| Field | What to record | Why it matters |
|---|---|---|
| Question and context | Exact wording, preceding turns, location, language, and relevant preferences | A changed budget or conversation can change the suitable options. |
| Platform and date | Product, visible mode, date, and whether web search was used or observable | Compare similar experiences; mark unknown settings as unknown. |
| Business appearance | Mentioned, cited, recommended, or absent; keep each as a separate field | A source link does not necessarily endorse that business. |
| Sources and facts | Linked URLs, quoted claims, current business facts, and saved answer evidence | Connect an error to a source you can inspect. |
| Action | Issue type, responsible owner, exact correction, and next review date | Turn observations into work that can be completed. |
Use a new conversation for a comparable fresh check and retain relevant settings. If personalization cannot be controlled, record it as a limitation. Do not infer that an assistant used a particular source when it does not expose one. An answer can be wrong without revealing a correctable citation.
For a small repeated sample, report raw counts: “recommended in 3 of 10 comparable observations.” Avoid calling this market share. Your questions, dates, products, and sampling process define what the count can describe.
Compare your business and competitors on the same baseline
Repeat the same questions under comparable conditions, count valid observations, and report mentions, citations, and recommendations separately.
Choose the product, market, language, and customer task before collecting answers. Record the businesses actually recommended alongside yours, including another type of solution when it serves the same task. Compare them within the same saved answers. Testing your brand on a difficult question and a competitor on its brand name does not provide a useful comparison.
For an initial repeat check, you could run five fixed questions three times in each selected assistant, using fresh conversations and recording the context each time. This is a practical sampling plan, not a required sample size. Keep each assistant and mode in a separate report. Record unavailable or failed responses rather than treating them as a business being absent.
| Measure | Calculation | What it establishes |
|---|---|---|
| Completed observations | 12 usable answers from 15 attempted checks | Three failed checks remain in the log, outside the answer-based denominator. |
| Business mentioned | 7 of 12 answers; 58.3% | The answer named the business, including passing mentions. |
| Business recommended | 4 of 12 answers; 33.3% | The answer explicitly offered the business as a fit for the request. |
| Business website cited | 2 of 12 answers; 16.7% | The answer linked the business's own site as a source. |
| One named competitor recommended | 6 of the same 12 answers; 50% | A comparison within this sample; several businesses can appear together. |
The categories overlap, so do not add the percentages. A recommendation may cite a third-party profile rather than the recommended business's site. Check the linked page before classifying the source. These observations describe the chosen sample; they are not market share, a ranking guarantee, or evidence that a page change caused a later result.
Turn a source gap into a correction someone can own
For each relevant source, record its URL, the claim it supports, whether the fact is current, and who can change it. Correct your own service page directly. Send a documented correction to an outdated profile. When a current independent comparison names a competitor, inspect the evidence and fit before assuming your business belongs there. Preserve editorial independence; do not treat placement as something an audit guarantees.
Prioritize facts that can mislead a customer, broken access or action paths, and missing evidence for a suitable offer. Give each issue an owner and an acceptance check, such as “the published coverage and postcode lookup both reject the unsupported area.” Re-run the original question set after the correction, keeping source accuracy and recommendation frequency as separate results.
Separate access, evidence, fit, and action problems
Classify the first supported problem. Missing information calls for a different fix from a genuinely unsuitable offer.
| Observation | Check next | Useful response |
|---|---|---|
| The intended page is missing from Google | URL Inspection discovery, crawl, indexing, and canonical information | Resolve the identified access or indexing issue; do not assume a successful live test means indexed. |
| The answer lists an old $140 price | Linked profile, current service page, and booking total | Correct the outdated source you control and request corrections elsewhere. |
| The business is mentioned but the ZIP is unclear | Coverage statement and postcode lookup | Publish exact coverage and route the customer to the relevant service. |
| The request requires a total below $180; your total is $195 | All mandatory fees and the requested package | Use another suitable offer or stop targeting that request. |
| A customer clicks but cannot book | Selected service, eligibility, current slots, and confirmation path | Fix the handoff before buying more traffic. |
Google's URL Inspection separates its stored index information from a live URL test. Use it for Google's handling of your page. It does not tell you which sources every AI product can access. A business absent from one answer is not enough evidence to diagnose a technical block.
When your offer fits and the available information is accurate, the assistant may still choose other options. Log that observation without inventing a ranking explanation. Paid acquisition is a separate business decision; it does not repair an independent answer by purchasing its inclusion.
FAQ
Does one missing ChatGPT mention mean my website is blocked?
Not necessarily. A missing mention can come from the question, available evidence, competing options, conversation context, or an access issue. Check the relevant page and repeat comparable observations before deciding what caused it.
Should I test a brand question or an unbranded buying question?
Use both, and keep the results separate. A brand question shows how accurately an assistant describes your business. An unbranded buying question shows whether it includes your business among suitable options without being prompted by your name.
How many questions do I need for a useful first audit?
Start with five to ten questions covering different customer decisions. That's a manageable first audit, not a representative measure of market visibility. Add questions when they test something you haven't covered yet.
Can I use this worksheet across several assistants?
Absolutely. You can use the same worksheet across assistants. Record the platform, product mode, context, and date for each observation, and compare similar experiences before combining their results.
What is a good AI recommendation rate?
There's no single target for every business. Establish your own baseline, repeat the same questions, and compare results over time. Keep the offer, platform, market, and sampling method comparable, and show raw counts and failed checks alongside percentages.
Continue with your next task
Sources and further reference
- Google Search Console: URL Inspection tool
- Google Search Central: AI features and your website
- OpenAI: Shopping with ChatGPT Search
The worked examples are teaching materials. Apply the method to your actual offer, records, and campaign terms.
