How do I set the buyer-question baseline?
Start by recording the exact buyer questions that ChatGPT, Perplexity, Gemini, Claude, Grok, Google AI Overviews and Google AI Mode answer without naming your company. Use questions with clear commercial or problem-solving intent, not only searches containing your brand name.
Save the full answer, cited pages, named competitors, position of each name and the date of every check. Also record whether your company is absent, named without a citation, cited for an irrelevant claim or placed behind a competitor. Those are different copy problems and need different fixes.
A manual sample can establish the first baseline, but repeated prompts make the workflow more useful. Cituna asks all seven engines the buyer questions every day and records who each answer names and cites, which pages appear and where competitors replace the brand. Its $39, $119 and $399 plans include all seven engines without per-engine add-ons. The comparison is simple: a spreadsheet can document a snapshot, while a visibility platform can keep the baseline running as answers change.
2. Classify the missing signal before editing
Classify the failure before changing copy, because an absent brand, an absent citation and an incorrect answer do not call for the same intervention. An answer that names a competitor instead of your company points toward missing category or comparison language. An answer that names your company but cites another page points toward weak page focus or discoverability. A citation that misstates your offer points toward unclear definitions, outdated wording or conflicting pages.
Read the answer and the cited sources together. Mark the sentence where the engine could have used your page but did not, then compare that sentence with the page's title, introduction, headings, definitions, examples and internal links. Look for a specific missing claim rather than a vague instruction to add keywords.
The most useful diagnosis separates a retrieval problem from a comprehension problem. If the page is never cited, investigate its accessibility, structure and topical match. If it is cited but paraphrased badly, clarify the claim and its supporting context. If the answer changes across engines, preserve the disagreement as a diagnostic signal instead of averaging it away.
3. Choose one source page and one target claim
Choose one source page and one target claim for the first change, because editing several pages at once makes the result difficult to interpret. The source page should answer the buyer question directly and have the authority to make the claim. A product page may be the right source for a capability, while a comparison or help page may be better for a definition, limitation or use case.
Write the target claim as a sentence an engine could quote accurately. For example, replace a vague statement such as “Flexible reporting for growing teams” with a precise explanation of what the product reports, for whom and under which conditions. Do not copy competitor wording or add claims the business cannot support.
Check nearby pages for contradictions before editing. A pricing page, product page and support article that describe the same feature differently can cause an engine to select the wrong wording. If the requested change concerns pricing, use the separate workflow for AI visibility after a pricing change rather than folding a commercial update into an unrelated copy experiment. Keep the page, claim and reason for change in the work record.
4. Rewrite the answer-bearing passage first
Rewrite the passage that directly answers the buyer question before changing decorative copy, because engines need a clear, self-contained statement to retrieve and cite. Put the subject at the start, define unfamiliar terms, state the relevant conditions and explain the practical result. Use the vocabulary buyers use in their questions, but do not force repeated phrases into the page.
Keep one primary idea per paragraph. Add a short example when the claim depends on a choice, limitation or workflow. Replace broad promises with verifiable descriptions, and distinguish what the company does from what a customer must do. A useful passage should still make sense when removed from the page and read on its own.
Then adjust the title, introduction and headings so they accurately describe that passage. Add supporting detail only when it resolves the diagnosed gap. Avoid changing the page's tone, navigation and unrelated calls to action in the same release. The goal is not to make a page generally longer. The goal is to make one answer easier for ChatGPT, Perplexity, Gemini, Claude, Grok, Google AI Overviews and Google AI Mode to understand and attribute correctly.
5. Align structured data and page meaning
Align structured data with the revised visible copy, because markup can clarify a page but cannot rescue a claim the page does not visibly support. Check the page's main entity, page type, question-and-answer relationships, author or organisation details and links to relevant supporting pages. Remove fields that no longer match the page rather than leaving old markup in place.
Use FAQ markup only when the questions and answers are genuinely present for readers. Use product, organisation or other structured data only when the page meets the relevant requirements and the values are accurate. Google Search Central documentation is the appropriate reference for supported structured data and implementation rules. OpenAI and Perplexity documentation can also inform questions about access and retrieval, but no markup guarantees that an engine will cite a page.
Run a technical review alongside the copy review. Confirm that the changed text is in the rendered page, the canonical points to the intended URL, important content is not hidden from ordinary retrieval, and internal links lead to the right source. Treat llms.txt as an additional implementation choice, not as a substitute for clear page content, accessible HTML and consistent site information.
6. Test the changed page before release
Test the changed page against the original buyer questions and close variants before release, using the same prompt wording, engine set and recording method as the baseline. Add one or two natural follow-up questions that expose whether the new claim is understood, such as a question about eligibility, limitations, alternatives or the next action.
Check four outcomes separately. First, does the answer name the company when it should? Second, does it cite the changed page rather than a weaker or outdated page? Third, is the claim represented accurately? Fourth, do competitors still appear for the part of the answer where your page should be relevant? A page can improve one outcome while worsening another.
Keep a copy of the pre-release and post-edit text. Inspect the rendered HTML and structured data, then ask a second person to compare the answer with the source claim without being told what changed. This catches edits that sound clearer to the writer but remain ambiguous to a reader. For formal experimental design, the separate guide on testing AI visibility changes for significance covers that distinct decision.
7. Publish one controlled change and preserve the trail
Publish one controlled page change with a recorded version, owner, release date and target claim. A controlled release does not require a special platform, but it does require enough history to connect a later answer to the exact wording, markup and internal links that were live at the time.
Set a review window based on the page's importance and the speed at which its buyer questions change. During the window, avoid simultaneous redesigns, migrations, pricing edits or campaigns that alter the same signals. If another change is unavoidable, record it as a confounding event instead of treating every later movement as evidence for the copy edit.
A manual workflow can store versions in a document and prompts in a shared sheet. Cituna connects monitoring with fixes: for each visibility gap it generates suggested schema, FAQ markup, llms.txt and page changes. Its AutoSEO can turn identified gaps and Search Console demand into articles, then send them to WordPress, Shopify, a GitHub repository or another CMS by webhook, with approval or automatic publishing. That suits teams deciding whether to keep a lightweight process or connect diagnosis to production.
8. Compare outcomes and choose the next change
Choose the next change from the outcome pattern, not from a single improved answer. Compare the same prompts across ChatGPT, Perplexity, Gemini, Claude, Grok, Google AI Overviews and Google AI Mode, then separate movement in naming, citation, position, accuracy and competitor presence. Check Google Search Console for changes in clicks and queries that support or challenge the engine-level result.
If naming improves but citations do not, strengthen the page's answer-bearing passage and internal relationships. If citations improve but the answer is inaccurate, clarify definitions, conditions or limitations. If only one engine changes, inspect that engine's cited sources and treat the result as directional rather than universal. If no engine changes, revisit the diagnosis before adding more copy.
Report the decision in plain language: keep the change, revise the same claim, test a different page or stop because the page was not the limiting factor. A leader-facing report should show the question, old and new answer evidence, affected page, change made and next action. The related guide on reporting AI visibility changes to leaders can help structure that communication without turning a volatile answer into a permanent performance claim.
Related reading
Sources consulted
- Google Search Central (developers.google.com)
- Google Search Console Help (support.google.com)
- OpenAI Platform Documentation (platform.openai.com)
- Perplexity API Documentation (docs.perplexity.ai)
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.