The routine holds.
- You finish the agreed collection.
- The answer evidence remains accessible.
- The team turns findings into useful work.
- You are still learning which questions matter.
A small, repeatable panel can answer your first visibility questions. Upgrade when you can name the work software needs to take over.
Published by Gist, which makes Gist GEO.
Sources checked October 19, 2026.
planned response slots
You can track AI visibility yourself with a spreadsheet. Start there if you have a small set of buying questions, a limited number of AI surfaces, and someone who can collect and review the answers consistently.
Pay for a tool when the work you need to repeat exceeds the time you can reliably give it, or when clients need evidence and reporting that your sheet cannot provide efficiently.
The important word is consistently. Asking an assistant about your brand once tells you what happened in that conversation. A useful tracking process gives you comparable observations and a record you can revisit.
Explore the workbook structure, then download a blank observation-column template. Examples are invented; no measurement is collected here.
One row per exact question version. Freeze it for the reporting period.
A blank CSV header, generated locally. Use the full four-tab instructions below to complete the workbook.
A sensible first objective is: “When we ask these buying questions on these AI surfaces, how often does our brand appear, and which sources are cited?”
That is a narrower claim than “What percentage of our market sees us in AI?” A panel of prompts you choose does not reveal total audience exposure or how often buyers ask those questions.
Choose questions from actual sales objections, customer conversations, or the decisions your content needs to support. Keep the original private material out of public AI tools. Rewrite the question without customer names or confidential details.
For a project-management product, a small panel might include:
Keep branded and unbranded questions in different groups. Naming yourself in a prompt changes what you are measuring. A response to “What are the disadvantages of Acme?” should not raise your unbranded discovery score.
Give each question a permanent ID and version. If you change the wording, start a new version and mark the change in your reporting.
Write down exactly where each answer came from. “Google” is too broad if you mean AI Overviews, AI Mode, or Gemini. Google says AI Overviews and AI Mode can use different models and techniques, and AI Overviews do not appear for every search. Google Search Central explains the distinction.
For each surface, record the model label when visible, language, intended market, collection method, and whether web search was selected or observed. If the model is undisclosed, record “not shown.”
Use a fresh conversation for each planned response. Check personalization settings rather than assuming a fresh window removes every source of context. OpenAI’s documentation says saved memories can influence search-query rewriting and approximate location can influence local results. Its temporary-chat controls also distinguish personalized and unpersonalized sessions. ChatGPT search documentation and temporary-chat documentation describe those settings.
These controls make your process easier to repeat. They do not turn your account into a representative sample of everyone using the service.
Use four tabs. The structure below is enough to start without inventing a composite visibility score.
One row per question version:
Freeze this tab for each reporting period. Put newly discovered questions into the next period’s panel rather than quietly adding them halfway through a comparison.
One row per collection attempt:
A link to a temporary conversation may stop working. Save the answer text and cited URLs in an approved location with appropriate access for your team. Use screenshots to preserve details the text does not capture.
Group observations by question version, surface, and collection cycle. Show:
Calculate mention rate by dividing answers mentioning the brand by usable answers scored for mentions. Label it “mention rate within this panel.” Calculate recommendation and owned-citation rates the same way, using the scored answers for each measure.
When every usable answer is fully coded, all three measures share a denominator. If citation capture is incomplete or a recommendation remains unresolved, exclude that unknown from the relevant measure and show the unscored count beside the result. Never silently treat it as “no.”
These are your own operational definitions. A vendor’s “Share of Voice” or “Citation Rate” may use a different denominator, so do not treat the labels as interchangeable.
Record content changes, product updates, and any measurement changes. Include the date, affected URLs or question IDs, owner, and what you expect to learn.
A later increase in mentions is a reason to investigate. A before-and-after spreadsheet alone does not prove that your edit caused it.
Use plain coding rules that another teammate could apply.
A mention means the answer body names your brand or an approved, unambiguous alias. A brand name appearing only inside a source URL does not count as a body mention.
A recommendation means the answer explicitly endorses your brand as suitable for the need in the question. Being listed in a neutral comparison does not automatically qualify. Keep a short supporting excerpt for each “yes.”
An owned citation means the response explicitly cites a URL on your defined list of owned domains as a source. An incidental product hyperlink does not qualify. A citation to a review site discussing you belongs in a separate earned-source note.
If an answer discusses the category and names no brands, it is still a usable answer. Record “no” for mention and recommendation. If you cannot decide whether a sentence is an endorsement, flag it for review rather than stretching the definition.
Have a second person check a small sample before you prepare the first report. Resolve disagreements and update the rules for everyone.
Change the bookkeeping inputs to see how the mention denominator and collection availability answer different questions. This is an illustrative calculation.
10 / 28 answers scored for mentions
28 / 30 slots produced usable answers. Keep missing-result reasons separately.
Default example: two searches produced no AI answer; 28 answers were usable and fully scored. Unknown coding must not silently become a “no.”
For a manageable pilot, choose 10 questions and two surfaces. Collect three planned responses for each question on each surface during week one, then repeat the same design in week two.
That is 10 questions × 2 surfaces × 3 repeats × 2 cycles = 120 planned response slots.
This is a suggested workload, not a statistically sufficient sample size for every business. Spread collection across preselected times. Three consecutive responses in one session give you a narrow view of variation.
Set the retry rule before starting. For example: allow one retry after a technical failure, following the provider’s usage limits. Keep both attempts in the sheet, but count only the successful retry for that original slot. If it fails again, leave the slot missing.
Do not rerun a completed answer because it omitted your brand. Do not replace a refusal or an absent AI Overview with a different question. Those outcomes should remain visible.
A missing AI feature is different from an answer that excludes you. Suppose one surface has 30 planned slots in a cycle. Two searches produce no AI answer; the remaining 28 produce usable answers that you fully score, and 10 mention your brand.
Report “10 of 28 usable answers mentioned us; 28 of 30 planned slots produced usable answers.” The conditional mention rate is 35.7%. The two missing answers remain in the availability count. Recording both as brand losses, or hiding them altogether, would tell a different story.
Keep refusals and technical failures visible for the same reason. If missingness changes substantially, investigate it before comparing the headline rates.
Review the answers before the percentages. A small panel is particularly useful for spotting a repeated factual error, a question where competitors keep appearing, or a source that repeatedly gets cited.
Choose one actionable finding. Check the underlying source, decide who owns the response, and record the work in your change log. Keep collecting the unchanged panel afterward.
Show engines separately. Also preserve an unchanged core panel when you expand the question set. Otherwise, a new collection of easier questions can make your chart look better even if the original questions did not improve.
Time collection, coding, quality checks and reporting. Compare any tool with the job your team is actually doing.
Keep the manual process when you can finish the agreed collection, retain the evidence, and turn the findings into useful work without losing a day to reporting.
It is especially useful while you are still deciding which questions matter. Buying capacity before you understand the panel can leave you monitoring hundreds of prompts that nobody uses to make a decision.
Time the pilot. Include collection, coding, quality checks, and reporting. Your own measured workload is a better buying signal than a generic recommendation to upgrade at a particular number of prompts.
Consider a tool when you are skipping collection, losing historical evidence, repeatedly rebuilding client reports, or maintaining enough brands and surfaces that quality checks are slipping.
Use a simple cost comparison. As a hypothetical example, if manual work takes six hours a month and a tested tool-assisted workflow takes two, the saving is four hours. At an internal cost of $60 an hour, that is $240 a month of analyst capacity.
Compare manual and tool-assisted costs over the same planning period, such as your first year. Include the subscription, one-off setup, and remaining review work. Then check whether the tool actually supplies the history, coverage, and reporting you need. Automation does not decide which customer questions deserve attention.
If you are evaluating Gist GEO, use your manual panel as a demo brief. Gist documents repeated responses in Deep Analysis, lighter Quick Scans, and Opportunities with impact and effort labels, action steps, and PDF or Markdown exports. See how Gist GEO works.
Bring five questions and one recent report to a Gist GEO walkthrough. Ask to see the path from an individual answer to a finding your team could act on, and confirm the plan’s coverage and limits. That makes the decision about your actual workflow.
Start with the sheet if it fits. Upgrade when you can name the work you need the tool to take over, and verify that it does.
Published by Gist, which makes Gist GEO. Product documentation checked October 19, 2026.
A tool earns its place when it can reliably take over work you understand. It does not decide which customer questions deserve attention.
Choose which optional analytics and advertising Gist can use. You can change your choices in Privacy settings. Privacy policy