One AI Visibility Score vs Per-Platform Scores: What Blending Hides

A blended AI visibility score takes answers from several engines and reduces them to one number. It looks tidy. The problem is what goes into the average.

As of October 2026, ChatGPT, Claude, Perplexity, Gemini, Google AI Overviews and Google AI Mode are separate products. They draw on different indexes, run different models and follow different rules for when to search. Google itself says AI Overviews and AI Mode "may use different models and techniques" and can return different links, and those are two surfaces from the same company.

So a single score is not a measurement of one thing. It is a weighted mix of several different measurements, and the weights are a modelling choice made by whoever built the tool. That doesn't make the number useless. It means you should know it's a model before you act on it. If you want the background on why engines behave so differently, see whether what works on one AI engine works on the others.

To average scores, you need units that mean the same thing on each side. AI engines don't give you that.

• Claude cites text spans. Anthropic's web search tool returns a short quoted passage of up to 150 characters per citation.
• Perplexity cites whole sources by a numbered index, so one citation points at a full page.
• OpenAI separates the URLs ChatGPT retrieved from the URLs it cited, and its documentation says the retrieved set is often larger.
• Google AI Overviews often do not trigger at all, so many queries produce no citation to count.

A span citation, a whole-source citation and a no-show are not the same event. Adding them up produces a number, but not a quantity. Citation is also a noisy signal of influence: mechanistic research (arXiv:2606.28358) repaired most missed citations without changing the sources, which shows citing and using a source can come apart.

The practical upshot: compare an engine with itself over time, not with another engine.

Every engine decides per query whether to search the web. That decision changes what kind of answer you are scoring.

With search off, the model answers from what it learned in training, frozen at a knowledge cutoff. With search on, it pulls live pages, and your crawler access and index status matter. These are different stages with different causes. You can fix access. Nobody can change a model's training data this quarter.

The mix is not published. ChatGPT searches at the model's discretion. Claude.ai search is a user toggle with no published default. Perplexity's Agent API requires developers to opt in to web search. Published estimates of how often ChatGPT searches range from about 18% of conversations (Profound) to 80.5% of prompts (cloro), because they measure different things on different populations.

This is why any single AI visibility score, including ours, averages an unobserved mixture of retrieval-on and retrieval-off answers. A per-platform score doesn't remove that problem. It does keep it contained inside one engine, where you can at least read the citations and tell whether search fired.

OtterlyAI ran a public test in 2026: two new domains, each loaded with about 1,000 AI-generated posts, tracked across seven AI engines for about four months.

Google removed both sites from its index. Bing kept them. Here is how the citations split for capybara-world.com, from Otterly's published figures:

• Microsoft Copilot, which grounds on Bing, produced 2,718 of the 3,934 tracked citations, about 69%.
• Perplexity and Claude cited the sites too.
• ChatGPT, Gemini, AI Overviews and AI Mode almost never did.

Now average that. One engine cited the site thousands of times and four barely cited it at all. A blended score would land somewhere in the middle and describe none of them. Worse, it would hide the cause. The per-engine view points straight at it: Google's AI surfaces stopped citing pages that Google had deindexed, and Copilot kept citing pages Bing still held.

The test has limits. It is one pair of sites, on a low-stakes topic, run by a vendor that sells tracking. It also showed that Microsoft's own AI Performance report counted 42,817 citations for the same site, against 3,934 in Otterly's prompt-based tracker, so any prompt sample is only a sample. The lesson still holds: blending turned a clear, fixable signal into noise.

Neither column gives you a rank. Google's Search Console report for AI features shows impressions only, with no clicks and no position, so no position metric for Google AI can be checked against ground truth. Read more on why in there's no rank to track.

A dashboard that tracks several AI engines can be useful. Its value comes from the per-engine breakdown, not the headline number.

Use each view for the job it can do.

Use a blended score when

• You need one line for a board slide and you label it as a modelled summary.
• You are tracking a broad trend on one tool, with the same prompts and weights each time.
• You want a quick first read, like the snapshot the free checker on this site gives for up to three buyer questions.

Use per-platform scores when

• Anything looks wrong. Start with the engine that scores lowest.
• You are deciding what to fix. Each engine maps to a crawler and index you can check: OAI-SearchBot for ChatGPT search, PerplexityBot for Perplexity, Claude-SearchBot for Claude, Googlebot and Search Console for Google, and Bing Webmaster Tools for Copilot.
• You are comparing before and after a change, such as server-rendering content or fixing a firewall rule.

The next move

1. Ask your tool for the per-engine breakdown and the prompts behind it. If it can't show both, treat the score as unverifiable.
2. Check index status engine by engine. Google Search Console and Bing Webmaster Tools tell you different things.
3. Keep the prompts fixed so the numbers are comparable. Our guide to building a prompt set covers how.

For the full method, see LLM visibility tracking across models, and for what a headline number does and doesn't mean, read what the free AI visibility score means.