Your working files.
Open access. Downloading a file does not subscribe you or create a discovery request.
Begin with questions a buyer would ask.
Choose a service your firm can deliver and a decision that precedes an appropriate enquiry. Build a small question set around that decision: what the service means, when it is useful, how to compare providers and what a buyer should verify. Keep brand-led questions separate from questions that do not name your firm.
Write the exact wording in Prompt definitions. Give each prompt an ID and a version, with the buyer task recorded alongside it, so a later edit cannot silently change the question or blur a consultancy evaluation with a local legal-service decision. Put geography into the question when it belongs to the buyer's decision, rather than adding a city to every test.
The starter set in the workbook is material for your team to adapt, and it does not represent search volume or prompts observed in real conversations. Ask the service lead whether each question reflects the work you want before investing time in observations.
Record the surface and the testing context.
An answer belongs to a particular product and context. Record the platform, whether search or browsing was used, the model information visible to you, date and time, language, location assumptions and relevant personalization. If a detail is unavailable, write unknown. Do not infer a hidden model version from the style of the answer.
Keep an approved evidence record with the observation. A public source URL and a permitted screenshot reference can help a reviewer understand what happened, but transcripts may contain account details or private inputs. Store them according to your firm's rules. This workbook does not connect to an AI account or collect answers automatically.
ChatGPT search can provide web sources and citations. Record what the actual interface shows instead of assuming that every mode performs the same retrieval or that an answer without visible citations used no external information. Platform documentation supports the description of the product, not a claim about a test you have not run.
Source: Searching the web with ChatGPT
Distinguish the observation you found.
A mention names the firm. A citation links to a source page. A recommendation suggests a provider in the answer's context. Use separate observation fields for the mention, citation and recommendation, retaining the cited URL where available, so an answer containing one of those elements is not mistaken for evidence of all three.
Review factual accuracy before celebrating inclusion. Record a suspected factual issue separately from the source supporting a correction, since an answer can name your firm while describing the wrong service, location or team. An old third-party description can deserve investigation even when your own site is accurate.
Avoid a single universal AI ranking. Treat each repeat as another recorded observation under its stated conditions, preserving changes in context without counting the result as an independent buyer or evidence of market-wide reach. Compare like contexts and keep a changed prompt version out of a supposedly unchanged series.
Turn recurring gaps into source work.
Look for a pattern that suggests a useful investigation: an unclear service description, missing expertise, an inaccessible page or contradictory business facts. Confirm the underlying issue directly before assigning a fix. A missing mention in one answer alone does not establish a technical defect.
Google's guidance for its generative search features keeps established SEO foundations relevant and does not prescribe special AI schema. Other platforms have their own behavior. Make your source useful and accurate, then test the published page and continue observation without promising that a particular engine will select it.
Give each action an owner and a verification method. For an incorrect office location, that can mean checking authoritative business information and eligible profiles, and for a thin explanation of a specialist service, it may mean a reviewed page addition. Do not create false locations or train a third-party foundation model as a proposed shortcut.
Source: Google's guide to generative AI search features
Keep reporting categories separate.
Search Console's generative AI report provides impressions for supported Google search features when enough data exists. Microsoft's AI Performance reporting describes citation activity on supported surfaces. Their definitions and coverage differ. Neither should be renamed a universal recommendation rank.
Website referrals, where observable, show visits with their own attribution limits. Attach the source, period and metric definition to a figure before comparing it with the log, keeping enquiries and qualified opportunities in their own matched records with follow-up evidence. An answer seen by a researcher is not a visit or an enquiry.
The existing article above remains the historical Growth Friday update, and this practical section adds a working method with sources checked separately. It does not retrospectively claim that today's platform reporting existed when the original update was written.
Source: Search Console generative AI performance report
Source: Microsoft AI Performance dashboard
Illustrative example: recording a question without inventing an answer.
A fictional consultancy defines question C01: What should a finance leader check before choosing post-acquisition integration support? It records version 1, a service-evaluation task and a proposed national buying context. No platform has been queried in this demonstration.
The example observation marks platform, model, answer, mention, citation and recommendation as not tested, and the follow-up action is to obtain an authorized, documented observation. This is a completed example of an honest preparatory record, rather than a fictional response dressed as test evidence.
The starter set also includes a legal question about what to verify before contacting an employment-law firm, and an architecture question about how to assess experience on an occupied-building project. None includes a real client name, confidential facts or a claimed response.
Adapt the review to professional judgment.
A legal practice's source review might investigate jurisdiction or an outdated attorney profile, using prompts checked for confidentiality and excluding identifiable matter details from the testing process. Marketing staff can record the issue, while the authorized professional decides the appropriate statement.
An architecture studio may discover that a project is attributed to the wrong firm or that a collaborator's role has disappeared. Check the original public project record and permission to reuse images or descriptions. Preserve credit and project status before using the observation to plan new content.
Use the workbook to keep an inspectable record.
Complete Prompt definitions first. Use Observations for each actual repeat and Evidence links for the source or permitted record. Put confirmed work in Actions. The fictional example remains on a separate sheet, and the Prompt starters sheet contains proposed questions rather than answers.
Record a next review date only when someone can own it. Keep the original observations, note changed conditions and document what a correction verified. If a report is unavailable or an answer cannot be retained, record that limit. The tool remains useful when it prevents a confident conclusion from weak evidence.
Sources and review.
Platform sources were checked on 2026-10-28. Recommendations are Growth Friday's planning guidance. Fictional scenarios demonstrate the tools, and are not client results or observed platform responses. No named human review is claimed.
