Explaining AI Visibility Score Limits to Stakeholders

Any AI visibility score is a vendor's model of how often, and how accurately, a set of engines mention you for a set of prompts. It isn't ground truth, because no ground truth exists to compare it with.

As of October 2026, no AI engine publishes where your brand sits inside its answers. OpenAI's whole disclosure on ranking is that ChatGPT uses "multiple factors". Google's generative AI report in Search Console shows impressions, with no clicks, no click-through rate and no position.

That doesn't make your score useless. It makes it a model with known inputs and known blind spots. Treat it the way a finance team treats a forecast: useful for direction, explicit about assumptions.

The next move is to learn what goes into your particular number. Our guide on what AI visibility scores average breaks down the usual components, and how an AI visibility score is calculated covers the arithmetic.

1. The score averages engines and retrieval states

Retrieval is per query and discretionary on every platform. Google AI Overviews often don't trigger. ChatGPT decides per query whether to search. Claude.ai search is a user toggle with no published default.

Nobody knows the mix for your prompts. Published estimates for ChatGPT disagree widely: Profound reports about 18% of conversations trigger a web search, Semrush's US clickstream put it at 34.5% in February 2026, and cloro reported 80.5% on a 100-prompt set. They measure different things on different populations.

An answer given without retrieval comes from training data frozen at a cutoff. OpenAI's model docs list February 16, 2026 for the GPT-5.6 family and August 31, 2025 for the Instant path. Nothing you publish this quarter reaches those answers.

The engines themselves differ too. Google says AI Overviews and AI Mode "may use different models and techniques". A blended number hides all of this, which is why per-platform scores beat one blended score for decisions.

2. Citation units differ by platform

Adding these together adds unlike units. Citation also isn't the same as use. Mechanistic research (arXiv 2606.28358) repaired over 90% of missed citations by amplifying specific attention heads, without changing the sources. If someone asks why citation rate and citation share moved in opposite directions, point them to how citation rate and citation share differ.

3. Google exposes impressions only

Search Console's generative AI report has impressions by page, country, date and device. Google also states that third-party tools don't have access to its internal ranking data. So any "position in Google AI" figure has nothing to be validated against. Bing's AI Performance report goes further, but Microsoft calls its Citation Share "an observational metric" and "not a ranking system or a competitive scoreboard".

Use these steps the next time the number goes in a deck.

1. Name the inputs on the first slide. List the prompts, the engines, the run dates, and whether runs were logged out with memory off. ChatGPT can use saved memories when it rewrites a query, so logged-in runs aren't reproducible.
2. Show per-engine results before any blend. Put each engine on its own row. Show the blended figure last, labelled as the vendor's weighting.
3. Report the trend on a fixed prompt set. Changing prompts mid-series breaks comparability. Decide the set once, using how many prompts you need as a guide.
4. Separate counts from model. Label which figures are counts (answers that named you, wrong facts found) and which are weighted outputs.
5. Say how the engine's queries differ from yours. OpenAI says ChatGPT search rewrites a prompt into one or more targeted queries. Your prompts are a proxy for an input nobody outside Bing Webmaster Tools can see.
6. Add a first-party count where one exists. In Otterly's 2,000 AI blogs experiment, Microsoft's AI Performance report credited one site with 42,817 citations, against 3,934 logged by Otterly's prompt tracker. The windows may not match, but it shows any prompt set is a sample.
7. Close with what you control. End on actions, not on the number.

Run through this list before you show a score to a board, a client or a sales leader.

• The prompt set is the same as last period, or the change is disclosed.
• Each engine has its own figure, with run dates.
• The blended score is labelled as a model, with its weighting stated.
• No slide uses the words "rank" or "position" for an AI answer.
• No figure compares your score with a competitor's score from a different tool.
• Any change in the vendor's scoring formula is flagged, because old and new scores aren't comparable.
• Known model changes at each engine during the period are noted.
• Crawler access was checked this period: robots.txt per bot and firewall or CDN responses.
• Wrong facts found in answers are listed with the page that should correct them.
• One slide states what the team controls and what it doesn't.

Good

• Per-engine results on a fixed prompt set hold steady or rise across several runs.
• A specific wrong fact, such as an old price, stops appearing after you fix it on a fetchable page.
• Every AI crawler you want gets a 200 response, and robots.txt matches what the firewall actually does.
• Stakeholders ask "which engine moved and why", not "what's our number".

Bad

• A single blended number compared against a competitor leaderboard. Rankings of brands mostly reflect brand size, as covered in whether AI visibility leaderboards mean anything.
• A target like "reach 80 by Q2" with no named inputs.
• Panic over a zero on one engine before anyone checks access. A zero can mean a blocked crawler, a retrieval-off answer or a narrow prompt set, as explained in what a zero AI visibility score means.
• Celebrating a jump that came from a new prompt set.

Run the same prompts monthly. AI answers vary run to run, so weekly reads mostly show noise.

Re-brief stakeholders quarterly on what the score includes. People forget the caveats faster than the number.

Re-baseline, and say so, whenever one of these happens:

• An engine changes the model its app serves.
• You change the prompt set or the engines tracked.
• Your vendor changes its scoring formula.
• Your CDN or firewall settings change. From September 15, 2026, Cloudflare set Training and Agent crawlers to blocked by default on ad-displaying pages for new domains and existing free-tier customers, so a site can lose AI access without anyone editing it.
• You ship an access or rendering fix. Mark the date so the trend has a reference point.

The levers with evidence behind them remove blocks. They don't add magic.

• Rendering. Vercel's crawler study found ChatGPT and Claude crawlers fetch JavaScript but don't execute it. Key facts need to be in the HTML before JavaScript runs.
• Per-bot robots.txt. OAI-SearchBot controls ChatGPT search. GPTBot controls training only. OpenAI says each setting is independent.
• Network reachability. A firewall or CDN can turn away a crawler that robots.txt allows. The AI Crawler Checker requests a page as each AI crawler and shows which ones get blocked.
• Snippet and index controls. noindex, nosnippet, max-snippet:0 and data-nosnippet keep content out of AI answers, so check none are set by accident.
• Stated facts. A specific price or one-line definition on a fetchable page gives an engine a sentence it can quote.

What nobody controls: whether an engine searches for a given query, which sources it picks, and its training cutoff. A July 2026 survey of 45 GEO studies found no technique with a stable, cross-platform causal effect on discoverability. Say that plainly. It protects you from being judged on a number nobody can move directly.

Can we compare our score with a competitor's score from another tool? No. Each tool picks its own prompts, engines, weightings and citation units, so two scores from different tools measure different things. Even within one tool, comparisons are only fair on the same prompt set and run dates. If you need a competitive view, compare per-engine results for the same prompts in the same tool, and remember that larger brands are cited more often regardless of what they publish.

Why did our score drop when we changed nothing? Several things outside your site move scores. Engines change the models their apps serve. The share of answers that use live retrieval shifts. Vendors revise scoring formulas. CDN defaults change, as Cloudflare's did on September 15, 2026. Check access first, then note any model or formula change in the period, then look at which engine moved before concluding anything about your content.

Should we set a target AI visibility score for the team? Set targets on inputs instead. Good targets are things the team can finish: every wanted AI crawler gets a 200, key facts render in HTML without JavaScript, no accidental nosnippet tags, and each wrong fact found in answers is corrected on your own pages. Track the score as a lagging signal alongside those inputs, not as the goal.

Is there any official data on AI visibility we can show? Some, with limits. Google Search Console shows impressions for AI Overviews and AI Mode, with no clicks, click-through rate or position. Bing Webmaster Tools has an AI Performance report with citations and grounding queries, the actual queries the AI ran, but as of October 2026 no API. Both are useful context. Neither gives a rank.

Does writing our own prompts make the score more accurate? It makes the score more relevant, not more accurate. Your own prompts reflect your buyers better than generic ones. But engines rewrite prompts into their own search queries, so any prompt set is still a proxy for the real input. Keep the set fixed over time so trends stay comparable, and disclose it when you present the number.