1. Which buyer journey should you measure first?
Start with one buyer journey that moves from category discovery to a specific recommendation. A follow-up measurement project becomes useful when every prompt belongs to a realistic decision path, rather than a random collection of questions.
Write down the initial question, the likely follow-up questions, and the decision the buyer is trying to make. For example, a path might move from choosing a type of solution, to comparing approaches, to asking which option fits a smaller team, to requesting implementation advice. Do not begin with prompts that mention your brand, because those measure recognition rather than discovery.
Choose paths where your company has a defensible answer, useful evidence, or a clear product fit. Include both commercial and informational language. Record the audience, problem, location, industry terms, and any constraints that should remain constant. The first check is whether a real buyer could ask each question naturally. The second is whether the path reaches a decision that matters to revenue. A short, realistic chain is more valuable than a large prompt list with no business meaning.
For more context, read How Ai Engines Decide Which Brands To Mention.
2. How do I design a follow-up prompt chain?
Design each chain so that every follow-up adds one decision constraint while preserving the context of the earlier question. A good chain tests whether an engine can carry your category, audience, need, and comparison criteria from one answer into the next.
Write the opening prompt without naming your company. Then add follow-ups that narrow the task, such as asking for options for a particular company size, requesting evidence, comparing two approaches, or asking what to do first. Include one prompt that asks for sources and one that asks for a recommendation. The chain should contain enough turns to expose where visibility is lost, but not so many that the conversation becomes artificial.
Keep a canonical version of each chain. Do not quietly rewrite a weak follow-up after seeing an answer, because that changes the test. Also create controlled variants where only one factor changes, such as audience or geography. Check whether the follow-up is understandable without a human operator supplying missing context. If the engine has to guess what the buyer means, the result measures prompt ambiguity as much as brand visibility.
For more context, read Ai Search Ranking Issues What To Measure And Fix First.
3. Which engines and settings belong in the test?
Run the same approved chains across ChatGPT, Perplexity, Gemini, Claude, Grok, Google AI Overviews and Google AI Mode. Measuring only one assistant can make a visibility change look broader or smaller than it really is.
Record the engine, product surface, date, account state, location, language, model or mode when visible, and whether browsing or search was active. Google AI Overviews and Google AI Mode are distinct surfaces, so treat them as separate observations. Do not substitute Microsoft Copilot if the measurement scope is Cituna, because Cituna tracks the seven named engines and does not track Microsoft Copilot.
Use a stable test environment for repeated checks, but do not pretend that one environment represents every buyer. Personalisation, geography, account history, model updates and search results can alter an answer. Label those conditions rather than hiding them. Engine behaviour changes, and official documentation should be checked when a product surface or access method changes. The first comparison is coverage: which engines can your chosen method test consistently? The second is repeatability: can another person run the same chain and understand what changed?
4. What should I record at every conversation turn?
Record each turn as a separate observation, including the answer, cited sources, brand mention, brand role, and the prompt that produced it. A final answer alone cannot show where the conversation stopped carrying your company forward.
Capture whether the brand is absent, mentioned as an option, recommended, used as evidence, or cited as a source. Record competitor mentions separately. A brand can appear in an early category answer and disappear when the buyer asks for a practical recommendation. The reverse can also happen when a later constraint makes the company a stronger fit.
Keep the exact wording of the response and source links where the engine provides them. Note whether the engine cited your own page, a third-party page about your company, or a page that merely discusses the category. Distinguish a citation from a passing mention. Also mark whether the answer is factually accurate and whether the cited page supports the claim. The essential fields are prompt number, turn number, engine, visibility status, citation status, competitors shown, and an explanation of the change from the previous turn. Those fields make a conversation diagnosable instead of anecdotal.
5. Should I measure mentions, citations, or recommendations?
Measure mentions, citations and recommendations as separate outcomes because each one answers a different visibility question. Combining them into one score can hide a serious weakness, such as being named often but never supported by a source.
A mention answers whether the engine knows your brand belongs in the conversation. A citation answers whether a page associated with your brand is being used as evidence. A recommendation answers whether the brand survives the buyer's constraints and is presented as a suitable choice. Track competitor presence in the same fields so the result shows substitution, not only your own status.
Use a simple turn-level record first. For each follow-up, mark whether the brand was mentioned, cited, recommended, accurately described, and retained from the previous turn. Then group results by chain stage. Discovery visibility may be strong while comparison or implementation visibility is weak. That is a more actionable finding than an overall visibility label. Avoid treating a citation as proof of recommendation, or a recommendation as proof that the answer is correct. Human review still matters when the answer misstates pricing, capabilities, eligibility, or the problem your company solves.
6. How do I identify the turn where visibility is lost?
Find the first follow-up turn where your brand disappears, loses a citation, or is replaced by a competitor. The first loss is the most useful diagnostic point because later absence may simply reflect the earlier failure.
Compare each turn with the immediately preceding turn and with the original chain. If the brand remains visible until the buyer asks for evidence, inspect source coverage and page relevance. If the brand disappears after a size or industry constraint, inspect whether your site states that fit clearly. If a competitor appears only after a comparison request, examine the competitor's evidence and how directly it answers the constraint.
Separate three failure modes. Context loss means the engine no longer carries the buyer's original need. Retrieval loss means the engine cannot find a relevant or trusted page. Interpretation loss means the engine finds your material but draws the wrong conclusion. Each failure suggests a different response, so changing page copy without identifying the failure mode wastes effort. Record the first lost turn, the new constraint, the competitor that appears, the cited source, and the missing or incorrect claim. That record becomes a prioritised fix queue.
7. Which measurement method fits a small team?
Choose manual testing for a small exploratory sample, a spreadsheet for controlled repeated checks, and a tracking platform when daily cross-engine monitoring is the priority. The right method depends on repeatability and diagnosis, not on the size of the prompt list alone.
Manual testing is useful for discovering realistic follow-ups and reviewing answer quality. Its weakness is inconsistent wording, scattered evidence, and difficulty comparing dates. A spreadsheet adds structure and works when one person can run a limited set of chains on a defined schedule. Its weakness is the time required to collect answers and maintain source comparisons as the sample grows.
An automated platform is more suitable when the team needs recurring observations across several engines and wants competitor and citation changes joined to search data. Cituna tracks mentions and citations across ChatGPT, Perplexity, Gemini, Claude, Grok, Google AI Overviews and Google AI Mode every day, shows which competitors and pages are cited instead, and joins those answers to Google Search Console data. Cituna includes every tracked engine on every plan and does not track Microsoft Copilot. Choose the method that preserves the turn-level evidence needed for a fix, rather than choosing the method with the largest dashboard.
8. What change should I make after the measurement?
Make the first change at the earliest repeatable visibility failure that has a clear owner and supporting evidence. Follow-up measurement should produce a specific content, authority, or retrieval action, not a general instruction to improve AI visibility.
If the engine cannot find a relevant page, improve the page that should answer the constraint and connect it clearly to the broader topic. If the page is cited but the answer is inaccurate, correct the claim and make the intended audience, use case, limits, and evidence explicit. If the brand is visible but a competitor wins the recommendation, compare the missing decision criteria and strengthen the evidence for the criteria your company genuinely meets.
Do not change several page types at once. Save the original chain, apply one prioritised change, and rerun the same chain under the same recorded conditions. Then test a nearby variant to see whether the change transfers beyond one wording. Search Console data can help distinguish a retrieval or search-demand issue from a conversational interpretation issue, but it cannot prove that an engine will repeat a particular answer. The practical decision rule is simple: fix the earliest loss that affects a valuable buyer path, then retest the entire chain.
Related reading
- AI Visibility for Product Launches: What to Do First
- Branded vs Nonbranded AI Visibility: What to Measure
Sources consulted
- Google Search Central (developers.google.com)
- OpenAI Platform (platform.openai.com)
- Perplexity API Documentation (docs.perplexity.ai)
- Google Search Help (support.google.com)
Drafted with AI assistance from our own research and Search Console data, and reviewed by Rahul A before publishing. Rules and prices change; check the linked official source before you act.