Monthly AI Visibility Report Template: Per-Engine Sections and Metric Definitions

A blended AI visibility score can hide the single most useful fact in your data: which engine is doing the citing.

Otterly's 2026 experiment shows this clearly. Of the 3,934 citations Otterly tracked for one test site, 2,718 (69%) came from Microsoft Copilot. ChatGPT, Gemini, Google AI Overviews and AI Mode almost never cited it. Google had deindexed the site while Bing kept it, so Copilot, which grounds on Bing, kept citing. One averaged number would have reported a middling score and hidden all of that.

Engines also differ in ways an average can't absorb. Google says AI Overviews and AI Mode "may use different models and techniques" and can return different links. Citation units differ too: Claude cites text spans, while Perplexity cites whole sources. There is no shared denominator to average across.

The move: give ChatGPT, Claude, Gemini, Perplexity, Copilot, Google AI Overviews and Google AI Mode their own rows. Treat AI Overviews and AI Mode as two products, not one "Google AI" line.

Readers trust a report until they find one number that isn't what it claims to be. Put one of three labels next to every figure.

• Measured. A first-party or mechanical record: Search Console's generative AI impressions, Bing's AI Performance citations, your own server logs of crawler requests, the status code a crawler gets back, your robots.txt rules.
• Sampled. Answers to prompts you chose and ran. These are real observations of a small slice. In Otterly's test, Microsoft's own report counted 42,817 citations for the same site where Otterly's prompt tracker logged 3,934. The time windows may not match, so it isn't an exact ratio, but it shows that a prompt set sees only part of the picture.
• Modelled. Any composite score, including vendor scores. Someone chose the weights. Any single AI visibility score also averages answers given with and without live retrieval, and you usually can't see which was which.

Modelled numbers are useful for tracking direction on a fixed method. They are not measurements, and the report should say so plainly. If you need the difference between two common citation metrics, read citation rate vs citation share before you define yours.

Use the same seven sections every month, in this order. Access comes first because nothing downstream works without it.

Keep robots.txt rows bot-specific. OpenAI states each crawler setting is independent: blocking GPTBot stops training use, while OAI-SearchBot controls whether you appear in ChatGPT search.

1. Freeze the prompt set. Use the same prompts as last month so changes reflect the engines, not your questions. The prompt set template covers categories.
2. Run access checks first. Request key pages as each AI crawler and compare the result with robots.txt. A firewall or CDN can return 403 to a crawler that robots.txt allows. The AI Crawler Checker shows which crawlers your firewall turns away.
3. Pull Google's numbers. Export AI Overviews and AI Mode impressions from Search Console. This Search Console generative AI report playbook explains what the report does and doesn't show.
4. Export Bing's AI Performance report. Record citations, cited pages and grounding queries. See reading Bing's AI Performance report.
5. Run the prompts per engine. Use logged-out sessions with memory off. For each answer, record whether it showed sources, which tells you whether retrieval likely ran.
6. Log wrong facts. Write down each incorrect price, product name, location or claim, by engine.
7. Compute the composite last. Name the method and label it modelled.
8. Write the change log. List what you fixed, the date, and which measured row moved afterward. Say plainly when nothing moved.

• Every engine has its own row, and AI Overviews and AI Mode are separate.
• Every number carries a label: measured, sampled or modelled.
• Access results list each retrieval bot by name, with status codes.
• Training bots (GPTBot, ClaudeBot, Google-Extended) are reported apart from retrieval bots.
• Google and Bing index status are reported separately. One being healthy says nothing about the other.
• Prompt results note the date run, the session state and whether sources appeared.
• The prompt set matches last month, or the change is stated at the top.
• No rank or position metric appears for Google AI features.
• Wrong facts are listed individually, not only as a score.
• Any known reporting bug in the period is flagged next to the affected numbers.
• The change log says what was fixed and what you expect to check next month.

Good signs

• Every retrieval bot gets a 200 on your key pages, and robots.txt matches what the firewall does.
• Key pages are indexed in both Google and Bing.
• Per-engine movement has an explanation you can point to, such as a fix shipped or a crawler unblocked.
• The wrong-fact list shrinks, or the same facts now appear correctly on a page the engines can fetch.

Warning signs

• A 403 to OAI-SearchBot while robots.txt allows it. That's a silent visibility failure, and fixing it is a priority.
• Google AI impressions fall to near zero. Check index status first, then the Search Console inclusion control, which can be inherited from a parent property. Also check for reporting bugs: a Google-side bug depressed impressions in these reports from roughly 12 August 2026.
• One engine carries nearly all citations, as Copilot did in Otterly's test. Check the other engines' indexes and access.
• The composite rose while no measured row moved. That usually means sample noise or a method change.

A warning sign is a list of things to check, not a verdict. Nobody controls what the models say. You control whether they can reach, index and quote your pages.

Run the full report monthly. Engines change models and defaults often, and a monthly cadence catches drift without drowning readers in noise.

Some checks need a faster loop:

• Access: re-run after every deploy, CDN change or firewall rule change. As of September 2026, Cloudflare blocks Training and Agent crawlers by default on ad-displaying pages for new domains and free-tier customers, so a site can lose AI access without anyone editing it.
• Prompt set: review quarterly. Change it only between reports, and note the change.
• Accuracy: re-check a wrong fact the month after you fix it on your site.

For leadership, condense the per-engine rows into a one-page summary using reporting AI visibility to your CEO. If you want to add revenue, read whether you can attribute leads to AI search first.