Choose questions that reflect a prospect’s decision

Your first deliverable is a question panel, not a score. As a practical starting point—not a measurement standard—we suggest 12 unbranded questions, drawn from sales enquiries, support requests, and searches you already know about.

Divide them between problems to solve, comparisons, and choosing a solution. For each question, record its exact wording, intent, customer need, the page that could answer it, and why you are tracking it.

IDIllustrative question for workshop softwareIntentPage to review
Q01How can I schedule repairs in a small workshop?Solve a problemScheduling guide
Q02Spreadsheet or software for managing workshop quotes?Compare approachesComparison of methods
Q03Which software should I choose to manage a repair workshop?Choose a solutionProduct overview

Replace these examples with your prospects’ wording. Each question should inform an editorial decision. If no page addresses the need, record “no page identified” rather than forcing a connection to your homepage.

Keep wording unchanged for repeat checks. Give a local variation or reworded question a new ID; it should not silently replace the original question.

Create a separate group for questions that already contain your brand, such as “Can AtelierFlux turn a quote into an invoice?” These check product descriptions; they do not measure unprompted brand discovery. UpByWeb’s measurement guide likewise recommends starting with customer questions and warns that a one-off test captures only a sample.

Create a log that lets you retrieve every observation

Create one row per attempt, including failed attempts. We suggest two rounds during your first week, starting a new conversation for each attempt and keeping the context as consistent as possible. Twelve questions repeated twice means 24 planned attempts per engine, not 24 guaranteed answers.

Your log should let you repeat the test and reread what you counted.

CategoryData to save
QuestionID, exact wording, intent, branded or unbranded group
ConditionsService, displayed mode, date and time, language, location, login status, session context
CollectionStatus, answer text or screenshot, reference to the saved file
PresenceBrand named: yes/no; link to your domain: yes/no; exact URLs
ReviewOther brands, other cited sources, passage to verify, factual note

Use three distinct statuses:

  • AI answer observed: the answer is accessible and saved; you can inspect its mentions and links.
  • Successful check without an AI summary: the Google search was checked, but no AI Overview appeared.
  • Technical failure: you could not verify the result. The brand’s presence remains unknown.

An empty cell should not mean “no,” “unknown,” and “not yet reviewed” at the same time. Record mentions and links only after inspecting an observed answer. For Google, do not count a link in standard search results as a link in the AI summary.

Before calculating anything, check that every row has a status and that every counted answer points to a saved record. If you retry a failed attempt, retain the failure and identify the retry as a new attempt. For the initial panel results, use the two scheduled rounds; report extra retries separately so they do not silently change the sampling base.

Calculate answer availability, mentions, and links separately

Set your counting rules before collecting answers. In this proposed protocol, a mention means that the answer text names the company. A website link means that at least one URL in the answer points to your selected domain, even if the trading name does not appear. Define accepted name variations and included domains or subdomains in advance.

An answer can contain both a mention and a link. Count it only once for each metric, even if the brand appears three times or five links point to its website. UpByWeb distinguishes simple mentions, linked citations, and what an answer says about a company; our counting rules make those observations explicit without inferring commercial impact.

For each engine and period, show the counts first: planned attempts, failures, successful checks, and observed AI answers. Then calculate:

MetricCalculationWhat it describes
Answer availabilityObserved AI answers ÷ successful checksWhether an AI answer appeared during completed checks
Mention rateAnswers naming the brand ÷ observed AI answersPresence of the name among observed answers
Link rateAnswers containing a link to the domain ÷ observed AI answersPresence of a link among those same answers

For Google, searches without an AI Overview remain in the availability denominator, but not in the denominator for mentions among observed summaries. Technical failures enter none of these denominators; report them separately. If no AI answer is observed, mention and link rates are not calculable, rather than zero.

Do not add mention and link rates together: they overlap. Do not combine engines into a headline rate either. A rate from a hand-picked panel describes that panel; it is neither market share nor an estimate of the audience exposed.

Worked example: AtelierFlux runs 24 attempts per engine

Entirely hypothetical scenario. AtelierFlux sells management software for repair workshops. During 2–8 November 2026, its marketing manager tracks 12 unbranded questions: four about scheduling, four about quotes, and four about choosing software.

She runs two rounds, on 2 and 6 November, in French with a France location setting. She preserves the wording, starts a new conversation for each attempt, and records the service, displayed mode, time, and login context. The base is 12 × 2 = 24 planned attempts per engine.

MeasureChatGPTGoogle AI Overviews
Planned attempts2424
Technical failures12
Successful checks2322
Observed AI answers2314
Successful checks without an AI summary08
Answer availability23/23 = 100%14/22 = 63.6%
Answers naming AtelierFlux6/23 = 26.1%3/14 = 21.4%
Answers linking to its domain4/23 = 17.4%2/14 = 14.3%

On ChatGPT, three of the four answers containing a link also name AtelierFlux. The six mentions and four links therefore cover seven distinct answers, not ten. On Google, the eight searches without a summary are valid observations; the two failures remain unknown. No overall rate combines the engines.

The manager also tests two questions explicitly naming AtelierFlux, each repeated twice: four additional attempts per engine. She keeps them in the factual-check group, outside the discovery rates above.

Reading the saved ChatGPT answers, she finds two inaccurate descriptions of the quoting feature. She checks current product documentation and notices that the public page does not clearly distinguish quoting from invoicing. This gap warrants a correction that helps readers; it does not prove that it caused the AI errors.

Her decision is not to commission twelve new articles. She assigns factual validation to the product manager by 10 November and the page update to the writer by 12 November. The deliverable includes a clear explanation of both features, a sample quote, and a dated reference. A check on 13 November verifies that the published page displays the approved text and working links. Completion depends on the actual correction, not on obtaining a mention.

Turn answer reviews into verifiable editorial tasks

A low score alone does not tell you which page to improve. Reread the answers while separating factual accuracy, explicit recommendation, simple brand presence, and links. A brand named in a list is not necessarily recommended; a positive description may contain an error.

For each suspected problem, copy the passage, find a current reference, and inspect the relevant public page. Examine the cited sources too: do they answer the question more precisely, explain a method, or provide evidence missing from your page? The purpose is to understand the reader’s need, not copy a source in the hope of being cited.

Choose a follow-up action based on what you can establish:

  • Correct misleading information when the fact can be verified and your page is ambiguous or outdated.
  • Expand an existing page when it covers the subject but misses the prospect’s specific situation.
  • Create a useful page when a distinct need has no satisfactory answer on your website.
  • Investigate further when the observation is isolated, contradictory, or insufficiently documented.

For each task, record the owner, deliverable, deadline, and expected check. “Improve AI visibility” is not verifiable. “Add a validated sample quote to the product page, then check its display on 13 November” is.

Prioritize the risk of misleading a customer even when the mention rate looks satisfactory. Conversely, absence from a single answer does not automatically justify a new page.

Compare two periods without manufacturing a trend

We suggest reviewing factual errors weekly and comparing the fixed panel monthly. Repeat the same wording on the same service, in the same language and market, under conditions as similar as possible. Record changes in mode, login status, or context.

Compare by engine, then by intent and question. Present counts before percentages: they reveal whether a change rests on one answer or several. Edikka’s published comparison documents differences in modes and sessions that limit interpretation of its observations over time.

For AtelierFlux, the next comparison keeps the 12 questions and two passes. If collection fails or fewer Google searches display a summary, the manager reports:

  1. Attempts, failures, and answer availability for each period.
  2. Mentions and links among the observed answers in each period.
  3. A supplementary comparison of question–pass pairs for which an AI answer is available in both periods.

The third view does not replace the complete results. It helps inspect changes on a shared basis while the full table shows changing coverage. For example, three mentions among 14 summaries and three among ten summaries represent the same mention count but different denominators—not evidence by themselves of broader visibility.

A new question enters a separate exploratory group before joining a future panel version. Even if more answers name AtelierFlux after its page correction, the finding remains descriptive. Repetition does not make the panel representative of all conversations, and the sequence of dates does not demonstrate an effect from the updated page.

Keep visits and sign-ups in a separate view

Maintain three separate views: presence in answers, measured acquisition, and commercial actions. Edikka’s measurement guide explicitly separates mentions and linked citations from clicks and commercial results.

The first table contains observed mentions and links. In the second, record Google clicks and sessions identified in your measurement tools, with their source and period. In the third, track a defined event: a completed sign-up, submitted demo request, or another useful action. Keep the source and scope of each measure visible rather than treating the tables as a single attributable funnel.

An impression is a display, not a visit. A click is an interaction, not necessarily a session recorded by an analytics tool. A citation without a visit remains a citation; it does not become a conversion.

In the hypothetical example, two links in the panel’s Google summaries do not mean two clicks or two sign-ups. The observation log does not measure how many people saw those answers. Likewise, sessions identified as coming from an assistant should not be presented as all traffic influenced by AI answers.

Add publication and correction dates to your tracking log. They provide context but do not establish that a page update caused a change in clicks or sign-ups. Continuing observation may be a more useful decision than claiming a commercial result.

Where Grow Mentions can reduce recurring work

Once you have defined the questions worth following, automation can reduce repeated collection and editorial preparation. According to Grow Mentions’ product documentation, it tracks ChatGPT and Google AI Overviews daily or weekly and observes whether answers name your company or link to your website. AI tracking and content creation can be configured separately, so you can observe without starting any writing.

When a gap warrants editorial work, the same documentation describes identifying questions and preparing drafts with sources, using connected data and documented product information. You provide the supporting product facts and review the proposed content. For a problem like AtelierFlux’s, the practical benefit is moving from an observation to a draft you can examine, rather than stopping at a mention rate.

The documented publishing workflow sends validated content through your GitHub publishing process and checks the page afterwards. You can require approval of each article or authorize automatic publishing within your project’s approved rules. Automatic publishing requires a GitHub-managed website and a configured deployment workflow. Follow-up checks are scheduled for 7, 14, 28, and 56 days after publication.

A spreadsheet remains suitable for a limited initial assessment. Automation becomes relevant when repeating questions, preparing drafts, and verifying publication regularly consume time. It does not change what the observations establish: the pricing documentation explicitly states that no plan guarantees a mention or ranking.

To see how the tracking and draft-review workflow is organized, explore the Grow Mentions demo without creating an account. You can still run the measurement protocol in your own spreadsheet.

Sources and references

Documents cited in this article.

Explore the tracking demo.

View the charts for a fictional project and inspect a proposed update.

View the demo