Why did my AI visibility score drop?

99Visibility Research · Updated October 21, 2026 · 5 min read

Short answer

An AI visibility score usually drops for one of six reasons. An AI crawler got blocked, often by a CDN or firewall default. Something on your site changed. The engine changed its model or routing. Answers varied, as they always do. The run returned fewer answers. Or the tool changed its formula. Check them in that order. Only the first two are problems you can fix on your side.

Key takeaways

Check access first: a firewall or CDN can block an AI crawler that your robots.txt allows, and you won't see it from outside.
Site changes such as moving content behind JavaScript or an inherited noindex or nosnippet tag can remove you quietly.
Model, routing and retrieval changes on the engine's side move scores, and nobody outside the vendor controls them.
A drop on one platform from a full, comparable run is a signal; a blended drop from a partial run is often noise.
Compare scores only when the prompts, platforms and formula version are the same.

The full answer: six causes, checked in order

A visibility score is an average. It blends answers from several engines, across several prompts, some given with live web search and some without. Our guide on what AI visibility scores average covers that in detail. When the average falls, one of the inputs changed.

The six causes, in the order to check them:

A new crawler block. A robots.txt edit, firewall rule or CDN default now turns an AI crawler away.
A site change. Content moved behind JavaScript, or a page picked up a noindex or snippet control.
A model or routing change. The engine now serves a different model, cutoff or retrieval path.
Answer variability. The same prompt gives different answers on different days.
A partial run. Some platforms errored, so the score rests on fewer answers.
A formula change. The tool now calculates the score differently.

The order matters. The first two are in your control and can wipe out a platform completely. The middle two are outside your control. The last two are about the measurement, not your brand.

Access and site changes: the drops you can fix

Crawler blocks you didn't make

From 15 September 2026, Cloudflare blocks Training and Agent crawlers by default on pages that display ads. That applies to new domains, new sites on existing accounts and all existing free-tier customers. Search crawlers stay allowed by default. A site can lose some AI access without anyone editing anything.

Firewalls cause the same silent failure. OpenAI's help centre says eligibility for ChatGPT search needs two things: allowing OAI-SearchBot in robots.txt and making sure your host or CDN accepts traffic from OpenAI's published searchbot IP addresses. If robots.txt says allow but the firewall returns 403, you're out of ChatGPT search.

robots.txt rules are also per bot. OpenAI states that each setting is independent. Blocking GPTBot stops training use. Blocking OAI-SearchBot removes you from ChatGPT search answers. A blanket AI block often removes the second by accident:

User-agent: GPTBot
Disallow: /

User-agent: OAI-SearchBot
Allow: /

Site changes

Vercel's crawler study found that ChatGPT and Claude crawlers fetch JavaScript files but don't execute them. If a redesign moved your pricing or product names into client-side rendering, those crawlers no longer see them. Googlebot renders JavaScript, so this one doesn't apply to Google.

On Google, a page must be indexed and eligible for a snippet to appear in AI Overviews or AI Mode. A template that adds noindex, nosnippet, max-snippet:0 or data-nosnippet removes content from AI answers. Check your Google and Bing index status separately. One being healthy tells you nothing about the other.

Engine changes and normal variability

Model and routing changes

"ChatGPT" is not one knowledge state. As of October 2026, OpenAI documents a February 16, 2026 knowledge cutoff for the GPT-5.6 family and August 31, 2025 for the Instant path. Routing between them changes what the model knows without live search. Vendors update these models regularly.

Retrieval paths change too. Perplexity's Sonar Chat Completions retired on 27 September 2026, and web search is opt-in on its Agent API. A tracker that didn't migrate may be scoring answers with no web access. See what the Sonar retirement did to tracking.

Variability

Retrieval is per query and at the model's discretion on every major platform. ChatGPT rewrites your prompt into its own search queries. Google AI Overviews often don't trigger. The same prompt on two days can give a grounded answer once and an ungrounded one the next. Learning how to tell whether an answer used live web search separates these cases.

Measurement changes: partial runs and new formulas

A score is only as stable as the run behind it. If one platform timed out, the score now rests on fewer answers. A single flagged answer among two returned weighs far more than one among fifty. Ask your tool whether failed AI responses count against the score. They shouldn't count as clean answers or as misses.

Formulas change as well. A formula that subtracts penalty points with no denominator punishes you for asking more prompts: more answers means more flags, even at the same error rate. A tool that fixes this will shift every score. Scores from two formula versions aren't comparable.

Prompt sets matter in the same way. Adding prompts in a category where you're weak lowers the average without anything changing in the engines.

How to tell a real drop from noise

Work through these before you act:

Break the score down by platform. A blended score hides which engine moved. A drop on one platform points to that platform's crawler or model.
Confirm the run was complete. Count the answers that returned, not the ones requested.
Confirm the inputs matched. Same prompts, same platforms, same formula version.
Test access directly. Request a page as each AI crawler and compare the result with your robots.txt. The AI Crawler Checker shows which crawlers your firewall turns away.
Check your deploy log. Look for rendering, template or robots changes around the date of the drop.
Rerun before reacting. A drop that disappears on the next run with the same inputs was variability.

If access and the site are clean, the cause sits with the engine or the measurement. You can't control those. Keep fixing what you can: fetchable pages that state your facts plainly.

Frequently asked questions

Can my score drop if I changed nothing on my site?

Yes. A CDN can change its defaults, as Cloudflare did on 15 September 2026 for Training and Agent crawlers on ad-displaying pages. Engines also change models, cutoffs and retrieval paths without notice. And answers vary between runs because each engine decides per query whether to search the web. Check access first, then compare per-platform results from complete runs.

Does a drop in Search Console AI impressions mean the same thing?

Not necessarily. Search Console's generative AI report shows impressions only, for AI Overviews and AI Mode. Google also confirmed a bug that depressed impressions in its performance reports from around 12 to 13 August 2026. Read our page on [Search Console AI impressions drops](/questions/search-console-ai-impressions-drop) before treating that report as confirmation of a score drop.

Should I add schema or publish an llms.txt to recover?

No evidence supports either as a recovery step. Ahrefs' matched test found AI Overview citations fell 4.6% after pages added schema. Google says Search ignores llms.txt, and no other engine documents reading it. Spend the time on access instead: crawler permissions, firewall reachability, server-rendered content and snippet controls.

How many runs do I need before trusting a trend?

The evidence base gives no fixed number. A practical rule: trust a change only when it shows up on the same platform across repeated, complete runs with identical prompts and the same formula version. A one-off drop that reverses on the next run is most likely answer variability, not a change in how engines see you.

Sources

This page is AI-assisted. It was written from a verified evidence base and checked against our editorial policy. Engine behaviour changes often; facts are as of October 21, 2026.