Google killed the FAQ rich result. It didn’t stop using the schema. And neither do AI engines.

Share

Post

Pinterest

Messenger

The news everyone got half-right

In May 2026, Google removed FAQ rich results from search — the collapsible question-and-answer boxes that used to expand under listings. Within hours every SEO newsletter ran a variant of “strip FAQPage schema off your site.”

That’s wrong, and Google’s own language is the proof. Per Search Engine Land’s coverage, Google explicitly stated that “the change is a search-appearance change, not an algorithmic one” and that it would “continue to use FAQ structured data to better understand pages, even though it will no longer display the rich result” (Search Engine Land, May 2026). FAQPage as a Schema.org type is still valid. The markup stays useful.

The deeper point isn’t about FAQ schema specifically. It’s about the assumption underneath the panic: if I can’t see it, the engine isn’t using it. That’s the same assumption that breaks 80% of the AI visibility advice circulating right now. So let’s unpack what AI engines actually do.

 

AI engines retrieve. They don’t rank.

The first thing to internalise — and what most “GEO” pitches gloss over — is that large language models don’t rank pages. They retrieve passages.

When a user asks ChatGPT “who do you recommend for X”, the model doesn’t return a top-10 list of URLs. It pulls chunks of text from its retrieval index, ranks those chunks by semantic similarity to the query, synthesises them with whatever it already encoded during training, and generates a new answer. The pipeline looks more like retrieve → re-rank → synthesise than crawl → index → rank → display (Discovered Labs, 2026).

This shift has three concrete consequences:

1. Page-level rank is the wrong unit. A page can be the #3 Google result and never get cited by Claude, because the chunks AI retrievers grab don’t depend on whole-page authority — they depend on whether a specific passage matches a specific query well enough to make it into the prompt context.

2. The retrieval index isn’t Google. Anthropic confirmed in March 2025 that Claude pulls real-time web context from Brave Search, not Google (TechCrunch). Brave maintains an independent index of 30+ billion pages with its own crawler and its own ranking signals — relevance, recency, link structure, domain authority, page layout (Brave Learn). A brand that dominates Google for “best dental clinic Dubai” can still be invisible to a Claude user asking the same question, because Brave hasn’t indexed the same set or weighted it the same way.

3. Different models pull from different places. Research from passionfruit.ai found that “only 11% of domains are cited by both ChatGPT and Perplexity” (Passionfruit, 2026). Same query, different sources. Same brand, different visibility. The single-number “AI visibility score” sold by many tools averages this gap away and tells you nothing useful.

What AI engines actually look at

When the retrieval and ranking research is read together, three signals come out repeatedly:

Brand search volume. Per the passionfruit data, brand search volume is the strongest single predictor of AI citations, with a 0.334 correlation. People searching your brand name correlates more strongly with AI engines mentioning you than any technical SEO factor does. This is also the hardest signal to fake — it shows up in Google Trends, autocomplete, and the kind of training data that propagates through model retraining cycles.

Referring domain count. Sites with 32,000+ referring domains are 3.5x more likely to be cited than those with under 200 (same source). The authority signals that always mattered in classic SEO matter more in AI search, not less — because AI retrieval reuses the same web indices that authority signals influence.

Mention density across authoritative sources. Brand mentions across authoritative sources correlate 3:1 over backlinks for AI citation likelihood (Passionfruit, 2026). A mention of your brand in a Reddit thread, a YouTube transcript, a Wikipedia entry, or a vertical-publication listicle weighs more for AI citation than a high-authority backlink would for Google ranking.

What’s notably absent from the “stuff that moves the citation needle” research: any direct attribution to a specific schema markup pattern. That doesn’t mean schema is useless — see below — but it does mean the “add FAQ schema and watch the citations roll in” pitch is a leap.

Where the schema debate actually lands

Pedro Dias — one of the more careful voices on this — frames it precisely. In The Inference, he writes: “no causal chain runs between ‘I added FAQ schema’ and ‘the model cited my page’ — what runs between them is a probability distribution, and the things you control affect that distribution in ways nobody can cleanly attribute, not even the people who created these systems” (Pedro Dias, 2026).

He’s right that there’s no clean causal chain. He’s also right that schema isn’t a magic input. The mistake is jumping from “no causal chain” to “schema doesn’t matter.” Schema isn’t a ranking input — it’s a clarity input. It helps Brave’s crawler, Bing’s crawler, and Google’s crawler parse what your page is about, which feeds the indices AI engines retrieve from. The deprecation of FAQ rich results doesn’t change that — Google said so explicitly.

The honest position is the one we argued in May: schema sits in Layer 1 of a four-layer visibility stack (entity foundation → corpus presence → answer shape → brand gravity). Skip Layer 1 and nothing above it holds. But there’s no schema markup that fakes Layer 4 either. Anyone selling AI visibility primarily as a schema fix is solving a Layer 1 problem and pretending it’s a Layer 4 outcome.

How to know where you actually sit

The framework above tells you what to look at. To know where your brand actually sits across the four major AI engines, the cheapest path is to ask them.

We built a free tool that does this systematically. It sends ten buyer-intent questions to Claude, ChatGPT, Gemini, and DeepSeek — across four categories (broad industry, niche use case, brand-named, and specialty/differentiator) — and reports each engine’s score separately, not as a blended average. It shows every prompt sent, every response returned in full, the specific competitors named in your place, and a parallel crawler audit covering ClaudeBot, GPTBot, PerplexityBot, and Google-Extended.

The methodology choices are deliberate and they reflect what the retrieval research above implies:

• Per-engine, not blended. Because the 11% overlap finding means the single number lies.
• Exact-string brand matching. Because LLMs can’t reliably judge their own outputs at temperature 0 — recent research shows substantial drift even at temperature zero due to floating-point non-determinism on hosted endpoints (arXiv, 2026). Substring matching gives you a deterministic score; LLM-as-judge gives you a number that changes between Tuesday and Thursday.
• Crawler audit included. Because Anthropic, OpenAI, Google and Perplexity each run multiple separate bots — training, search-indexing, real-time retrieval — and many B2B sites unknowingly block one of them at the CDN layer.

The output isn’t a guarantee. Nobody can guarantee citation outcomes for the same reason Pedro Dias points out: the chain is probabilistic, not causal. What the tool gives you is a defensible read of where you stand today across the four engines that account for roughly 95% of consumer AI chat traffic.

The summary in two lines

Treat “deprecated” and “ignored” as different words — they almost always are. Treat “AI visibility score” and “ranking in Google” as different problems — they almost always are too.

Run the check on your own brand. Three checks per email per 30 days, no signup, no upsell, full per-engine breakdown.

Source: Foreground Digital

(Visited 1 times, 1 visits today)

Share

Post

Pinterest

Messenger

Tags: ai search, brand visibility, geo, llm, retrieval, schema, tools Last modified: May 15, 2026