AI Visibility Scores: What They Average and How to Read One

An AI visibility score is a single number a tool computes from a sample of AI answers. The tool picks a set of buyer prompts, sends them to several engines, and records what comes back. It then counts things like whether your brand was named, whether your site was cited, and whether the facts were right. Finally it weights those counts and rolls them into one figure, often 0 to 100.

Every step involves a choice. The vendor chooses the prompts, the engines, the components, the weights and the curve. Those choices are reasonable or unreasonable, but they are not discovered from data. Two honest tools can make different choices and give the same brand different scores, which is why AI visibility tools disagree so often.

This applies to the 0 to 100 snapshot from the free checker on this site too. It scores up to three buyer questions with the 99Visibility score. That makes it a useful snapshot of those questions on those engines, not a universal measure of how AI sees your brand. For the component-by-component mechanics, see how an AI visibility score is calculated.

What this means for you: treat any score as a summary of one tool's sample. The useful questions are about the sample, not the number.

A blended score assumes a citation on one engine equals a citation on another. It doesn't. Engines cite different units, search the web under different rules, and expose different amounts of what they did.

Anthropic's web search documentation describes citations as short quoted spans. Perplexity points to whole documents. Adding one to the other is like adding sentences to books. Google's AI features page states that AI Overviews and AI Mode may use different models and techniques, so even two Google surfaces aren't one channel.

There's a second gap: cited is not the same as used. OpenAI's own documentation says the number of sources a model consulted is often greater than the number it cites. A score built on citations measures what got credited, not what influenced the answer.

What this means for you: any cross-engine total is the vendor's model of equivalence. Ask how it converts between units. If the answer is "it doesn't", the blend is mostly weighting.

AI answers come from two places. Some come from the model's training, frozen at a knowledge cutoff. Others come from a live web search at answer time. Most consumer apps decide per query which to use, and the observer often can't tell.

That matters because the two kinds of answer respond to different things:

• Parametric answers (no search) reflect what the model learned before its cutoff. OpenAI's model docs, checked in August 2026, list a February 16, 2026 cutoff for the GPT-5.6 family and August 31, 2025 for the Instant path ChatGPT uses. Nothing you publish this quarter reaches these answers.
• Retrieval answers (with search) depend on whether a crawler could reach your page, whether it was indexed, and whether it entered the model's context for that query.

A score that averages both is averaging a number you can't move this quarter with a number you partly can. If the mix shifts (an engine starts searching more, or less), your score moves even though nothing about your brand changed. That is one common reason a score drops without any edit on your side.

What this means for you: prefer tools that record whether each answer searched the web, and read the searched and unsearched answers separately when they do.

Before you act on a number, get answers to five questions.

1. Which prompts? Real retrieval queries are generated by the engine and aren't visible. OpenAI says ChatGPT search rewrites your query into one or more targeted queries. A prompt set is a proxy for an input nobody outside the engine sees.
2. Which engines, and which models? "ChatGPT" isn't one thing. Plan tier and routing change the model, and an API call may not match what the app serves.
3. Was search on? If the tool can't say, assume the score mixes both states.
4. What are the components and weights? Mentions, citations, accuracy and sentiment answer different questions. Know which one moved.
5. Did the formula change? Scores from before and after a formula revision aren't comparable. Good tools record which version produced each result.

Also ignore any claim of an AI "rank" or average position. Google's generative AI report in Search Console shows impressions only, with no clicks, click-through rate or position, so there's nothing to check a position against.

What this means for you: a score is readable only alongside its method. If the method isn't disclosed, use the number for trend within that one tool, and nothing else.

A blended score is fine when engines move together. It misleads when one engine behaves differently from the rest, which happens often because each engine draws on a different index.

Otterly's 2026 experiment shows how large the gap can be. Of the 3,934 AI citations Otterly tracked for one test site, 2,718 came from Microsoft Copilot, while ChatGPT, Gemini, AI Overviews and AI Mode almost never cited it. Google had deindexed the site; Bing had not, and Copilot grounds on Bing. A single blended number would have shown modest visibility and hidden both facts: strong on one engine, near zero on four.

The same experiment shows what any prompt sample is. Microsoft's own AI Performance report counted 42,817 citations for that site, more than ten times Otterly's prompt-based count, though the time windows may not match. Where an engine publishes its own count, that count sees more than a sample does.

Use per-engine numbers first when:

• Your buyers concentrate on one or two engines.
• The blend moved and you need to know where.
• You suspect an access or indexing problem, which usually hits one engine's pipeline, not all of them.

The trade-offs are laid out in one score versus per-platform scores.

The evidence is plain about where the levers are. A July 2026 critical survey of 45 GEO studies (arXiv 2607.14035) found that already-retrieved content can change how it's cited, but no reviewed technique shows a stable, cross-platform causal effect on being discovered in the first place.

In your control (access):

• Server-side rendering of the content that matters. Vercel's crawler study found ChatGPT and Claude crawlers fetch JavaScript files but don't execute them, so client-rendered content may be invisible to them.
• Per-bot robots.txt rules. OAI-SearchBot controls ChatGPT search; GPTBot controls training only.
• Firewall and CDN reachability. A CDN can block a crawler that robots.txt allows.
• Index and snippet eligibility: noindex, nosnippet, max-snippet and data-nosnippet keep content out of AI answers.
• Specific, correct facts (prices, numbers, one-line definitions) stated on pages that can be fetched.

Not in your control:

• Parametric answers before the next model generation.
• Whether an engine decides to search for a given query.
• Your position in the model's context window.
• Brand size. Larger brands dominate citations, which is why industry benchmarks are only meaningful between brands of similar size.

Nobody can promise a score. What you can do is remove the blocks.

Work from the gates outward. Each step either removes a block or tells you a block isn't there.

1. Check crawler access. Confirm robots.txt allows OAI-SearchBot, PerplexityBot and Claude-SearchBot if you want to appear in those engines' answers. Then confirm your firewall actually serves them. 99Visibility's free AI Crawler Checker requests a page as each AI crawler and shows which ones get turned away.
2. Check rendering. Compare raw HTML with the rendered page. Anything that appears only after JavaScript runs, such as pricing or product names, is a fix target.
3. Check index status engine by engine. Look at Google Search Console and Bing Webmaster Tools separately. One being healthy tells you nothing about the other.
4. Check accuracy. Read what each engine says about your prices, products and company. Fix contradictions on your own site first.
5. Then read the score. Use the per-engine numbers to find where the gap is, and the blend only as a trend line within one tool.

When you report upward, say which numbers are measurements and which are the tool's model. The playbook on explaining score limits to stakeholders covers how.

• Comparing scores across tools. Different prompts, engines and weights make the numbers incomparable.
• Reading a drop as a content problem. Check access, indexing and retrieval mix first.
• Chasing the blend instead of the engine. A small rise in the average can hide a collapse on the engine your buyers use.
• Buying tactics to raise the number. Google says structured data isn't required for generative AI search, and Ahrefs' matched test found AI Overview citations fell 4.6% after pages added schema. Google also says Search ignores llms.txt. None of these is a lever for AI answers.
• Treating citations as influence. Engines consult more sources than they cite.
• Comparing across a formula change. If the vendor revised its scoring, old and new numbers measure different things.

Can I compare AI visibility scores from two different tools? Not directly. Each tool picks its own prompts, engines, models, components and weights, so two tools can score the same brand very differently and both be internally consistent. Compare each tool against itself over time, using the same prompt set. If you need a cross-tool view, compare per-engine findings, such as whether a specific engine named you for a specific question, rather than the blended totals.

Does Google publish an AI visibility score? No. Google's generative AI report in Search Console shows impressions for AI Overviews and AI Mode, with no clicks, no click-through rate and no position. Google also says third-party tools don't have access to its internal ranking data. Any Google AI rank or position you see in a tool is that tool's estimate, with no first-party figure to check it against.

Why does the same prompt give different answers on different days? Engines decide per query whether to search the web, rewrite prompts into their own queries, and update models without notice. Logged-in sessions can add memory and custom instructions. Each of these changes the answer without anything changing on your site. For repeatable measurement, run prompts logged out with memory off, and look at trends across many prompts rather than any single answer.

Can schema markup or llms.txt raise my score? The evidence doesn't support it. Google states structured data isn't required for generative AI search, and Ahrefs' matched-control study found AI Overview citations fell 4.6% on pages that added schema. Google says Search ignores llms.txt, and no other engine documents reading it. Existing markup does no harm, but it isn't a lever. Access fixes are where the evidence is strongest.

Is a per-engine score more accurate than a blended one? It's more honest, not more precise. A per-engine score still depends on the tool's prompts and still mixes searched and unsearched answers. What it avoids is adding incompatible citation units across engines and hiding one engine's behavior inside an average. When engines diverge sharply, as they did in Otterly's experiment, per-engine numbers show the gap that a blend would bury.