AI Overview citation: 30-day field report, Part 2 — Google says FAQ schema isn't required. We keep getting cited.

On May 15, 2026, Google published its first official guide to optimizing for generative AI search — and the headline finding contradicted most of what AEO consultancies have been selling. FAQPage JSON-LD, content chunking, llms.txt files, AI-specific rewriting: none of it required. Three days later, our pillar on what a creative technology agency does showed its strongest AI Overview citation density yet — five inline citations across six sections, with no re-crawl since April 24. The schema isn't doing the work. What's doing the work is the answer-first prose underneath the schema, the clear semantic sectioning, the named-entity authority, and the fact that AI Overview synthesis recomputes over the indexed corpus on a near-weekly cycle independent of when Google last crawled the page. The discipline isn't in the markup. It's in what you put on the page before you mark it up.

The weekend I changed my mind

I spent most of Friday afternoon reading Google's new guide to optimizing for generative AI search. It was published that morning — May 15, 2026 — and a lot of people in the SEO world were already in the comments arguing about what it meant. I had a particular reason to care. The thing Google said most clearly in the guide was that the entire toolkit a swath of "AEO/GEO" consultants have been selling — FAQPage JSON-LD as a magic schema, content chunking, llms.txt files, rewriting your prose specifically for AI systems — isn't required. Not "is harmful." Not "is deprecated." Just: you don't actually need it.

I had two reactions in sequence. The first was satisfaction — we'd never bought the chunking-and-llms.txt story, and the guide validated the skepticism. The second was more uncomfortable. We had bought the FAQPage JSON-LD story. It was in the format pattern we'd documented and published, and one of the first things I'd recommend to a client. If the schema isn't required, what's actually doing the work?

I left the question open over the weekend.

Monday morning I ran our usual rotation pull on the brand-pillar queries. One of them — the pillar that defines what a creative technology agency does — came back with the strongest citation density we've ever recorded on that query. Five inline citations across six different sections of the AI Overview, our domain in the right-panel references, and organic rank #2 underneath. We hadn't touched the page since April 24. We hadn't changed the schema. We hadn't republished. Google's crawler hadn't been back to the page in 24 days. The AI Overview just liked us more.

That's the thing I want to talk about here. The shape of "you don't need the schema to earn citations" lined up with the shape of our own data. And that has implications for what the operating system actually is — versus what we'd been telling ourselves it was.

This is Part 2 of the 30-day field report we published two weeks ago. Part 1 documented a format pattern we'd pulled out of the first wave of citations. This one is the rewrite of that pattern after the data and Google's own guide collided.

What the guide actually says, in plain English

The full thing is at developers.google.com/search/docs/fundamentals/ai-optimization-guide and the Search Relations team's blog post is at the equivalent path under /search/blog/. The mythbusting section is short. Google names five tactics and says, directly, that you don't need them:

• You don't need llms.txt files or any other "special" markup aimed at AI crawlers. Standard structured data is fine. There's no AI-specific schema.
• You don't need to chunk your content into small pieces for AI to understand it.
• You don't need to rewrite your prose for AI systems. AI handles synonyms and meaning the same way humans do.
• Seeking inauthentic brand mentions and citations doesn't help. Quality is the signal.
• Special structured data isn't required for AI search and there isn't a special schema.org markup to add.

The one nuance: Google still recommends using structured data as part of a normal SEO program, because structured data still earns you rich-result snippets in the regular SERP. The FAQPage JSON-LD still buys you the visual SERP treatment. It just isn't what's making AI Overview cite you.

Matt Southern at Search Engine Journal called the guide "the most explicit guidance yet on what you should and shouldn't do for generative AI features in Search." I agree. It's the first time Google has said the quiet part out loud.

The harder question is: if the schema isn't doing the work, what is?

Our pillar, in plain English

The page I mentioned at the top is /insights/what-creative-technology-agency-does. It went up April 24 with what I'd call a fairly disciplined AEO patch: an upfront answer block of about 160 words at the very top of the piece, a four-question FAQ section near the bottom, the FAQPage JSON-LD wrapped around the FAQ, a few institutional citations to McKinsey and Gartner for stat density, four internal links to related pillars, and a dozen or so external links to authoritative sources.

Here's what's happened since then. I'm going to walk it day by day because the rhythm is what matters:

• April 24: Pillar shipped with the AEO patch. Google crawls it that day. Last GSC crawl: April 24. (That date matters later.)
• May 8: First citation event. Our 30-day field report goes live, and inside it we document this pillar as Query 1 — multi-section AI Overview citations on what is a creative technology agency. We feel good. We tell ourselves the format pattern is working.
• May 11: Monday rotation pull. Zero citations on the same query. Wikipedia and a GitHub list-of-creative-tech-things entered the citation mix. We dropped out. I don't sleep great that night.
• May 13: A second rotation pull on the same query confirms the regression — still zero citations. We retract the "the trigger fired" call from May 8 and set a new rule: a citation surface only counts as durable when it holds across two consecutive cycles.
• May 15: Friday weekly deep dive. Our pillar is back at two inline AI Overview citations plus a right-panel reference. Recovery, but cautious.
• May 18 (this past Monday): Five-plus inline citations across six sections of the AI Overview, plus the right-panel reference, plus organic rank #2 on the SERP.

What did we ship between May 13 (zero citations) and May 18 (five-plus)? Nothing. We didn't touch the page. We didn't update the schema. We didn't republish. The GSC inspect_url verdict confirms the page's last crawl is still April 24. Google's crawler hadn't been back to look at the page during the entire citation-density growth window.

What changed on Google's side? On May 14, Google switched AI Overviews and AI Mode to Gemini 3 as the default model globally. The next day they published the AI Search guide. The citation density that grew on our pillar across that window wasn't driven by anything we did. It was driven by Gemini 3's recompute of the same indexed corpus we'd been sitting in since April 24.

The unsexy way to say this: AI Overview citation density is decoupled from when Google last crawled your page. The recompute runs over the existing index. You can grow a citation surface without republishing.

And the corollary, which is the thing Google said in the guide: when they tell you fresh dates and dated content aren't a separate signal for AI Overviews, they mean it. Our 24-day-stale pillar earning its best citation density yet is evidence on their side.

Was the schema doing anything at all?

This is the question I spent the weekend on, and I want to share both sides because I think the honest answer is "probably not, but I'm not 100% sure."

The first read of the data is that Google's guide is directionally right, and the FAQPage JSON-LD on our pillar was a coincidence — there but not load-bearing. Under this read, what was actually doing the work was the prose. The 160-word answer block at the top of the pillar. The clear H2/H3 sectioning. The McKinsey and Gartner citations giving the synthesis layer something authoritative to anchor on. The schema wrapped around the FAQ section was a side-effect of the AEO consultancy industry recommending it; we added it, we earned citations, the two correlated, but correlation isn't cause. The same way a high-quality article tends to have nofollow outbound links — but adding nofollow doesn't make your article high-quality.

The second read is more cautious. Maybe Google's "not required" is technically accurate but a little incomplete. Maybe what "not required" really means is "you don't have to have it to earn citations" — not "having it doesn't help." Under this read, the FAQPage JSON-LD might still be a small load-bearing signal that makes our content more parsable for the synthesis layer, even if Google doesn't require it.

The honest version: we can't distinguish these two reads from one pillar's data. We'd need a control group of pillars without FAQPage JSON-LD earning comparable citations, and we don't have that yet.

But — and this is the small detail that makes me lean toward the first read — Google's rich-result detector currently isn't surfacing any FAQ markup on our pillar at all. The inspect_url API doesn't report a FAQ rich-result for the page. So whatever's happening, the FAQ schema isn't being read by Google's rich-result systems right now, and the pillar is still earning peak citations. We're basically already running the "no detected FAQ schema" experiment in production, and the pillar is winning.

Reading 1 — "the schema is incidental, the prose is doing the work" — fits the data better than Reading 2.

What this means for our operating system

In Part 1, we documented seven format-pattern elements. After this week's data, two of them need updating. Here's the redline:

The FAQPage JSON-LD claim. What we said in Part 1: add the FAQPage JSON-LD as a load-bearing AEO format signal. What we're saying now: keep the FAQPage JSON-LD if you have it, because it still earns SERP rich-result eligibility, but it isn't load-bearing for AI Overview citation. What's load-bearing is the answer-first prose pattern the schema happens to wrap. The schema is the wrapper, not the engine.

The dated-freshness claim. What we said in Part 1: dated freshness within 30-60 days is a citation-surface signal. What we're saying now: dated freshness helps you rank in the standard SERP. It is not a separate AI Overview citation signal. Our 24-day-stale pillar hitting peak citation density on May 18 is direct evidence.

And we're adding one new element to the pattern:

Citation density is decoupled from crawl freshness. Re-publication isn't required to grow citation density. AI Overview synthesis recomputes over the indexed corpus on what looks like a near-weekly cycle. The implication for cadence is something like "publish more pillars, let the recompute find them" rather than "publish less and refresh more."

The other five elements from Part 1 hold up:

• The answer-first prose at the top of the pillar — the 134-to-180-word upfront answer block — is still load-bearing. The AI Overview citations on this pillar lift heavily from that exact passage. It earns the citation surface for the same reason a good lede earns the click.
• Stat density of about one named-source citation per 200-300 words — still load-bearing. The synthesis layer prefers corroborated claims, and our McKinsey and Gartner anchors give it what it wants.
• Internal-link mesh of four-to-nine links to related pillars — still load-bearing, though I want to be honest that we may have been over-weighting how much it matters relative to the prose.
• Semantic HTML with H2/H3 sectioning — load-bearing. The structure of the pillar is what lets Google surface "Roles You Might Find" and "Intelligent Systems" as discrete citation anchors inside the AI Overview, instead of just synthesizing a generic summary.
• Named-entity authority — load-bearing. The McKinsey, Gartner, and Bluegrass mentions make the pillar legible as a synthesizable source. Take those out and we're a no-name article making claims.

The shape of the operating system didn't change. The plumbing did.

The three counter-arguments I want to give air to

If I were reading this piece skeptically, here are the three things I'd push back on. I'll steel-man them and answer them.

"You got lucky on a low-volume query." It's true that what is a creative technology agency is a low-base query — DataForSEO's keyword_data returns no measurable monthly volume. The signal there is going to be noisier than on a 12,100-per-month query like what is an ai agency, where we still haven't earned citations. It's possible the low-base SERP just hasn't accumulated enough institutional competition to push us out.

Fair point, but: five-plus citations on a low-base query is still five-plus citations. And the citation count in the synthesis layer is largely decoupled from query volume — Gemini-3 cites whoever it thinks corroborates the synthesis, regardless of how many users are searching the term. The mechanism is the mechanism.

"Your competitors on that SERP are softer than the BCG/AWS-class corpus on next-door queries." Also true. The pages currently above us on the standard SERP for what is a creative technology agency are agencies and educational sites — Future Colossal, Bluegrass Digital, SMU Meadows, Indeed. Compare that to our what does an ai-powered agency do differently pillar, where the AI Overview cites BCG, Google Cloud, and AWS, and we're nowhere in it. Maybe the creative-technology pillar is just a softer SERP to break into.

Also fair, but here's the thing: this is exactly how content velocity compounds in a new cluster. You win the soft SERPs first. The Cluster 7 pillars we shipped on May 15 — the law-firm piece and the M&A due-diligence piece — face an equivalently soft early-stage SERP. The format pattern from this pillar is exactly what we have to test against on those. The behavior on this pillar is the proof-of-concept that informs how we publish into Cluster 7.

"24 days isn't enough to call durability." Also fair. We set the framework on May 11: a citation surface counts as durable only when it holds across two consecutive retest cycles. This pillar is now at cycle 3 — May 11 zero cited, May 15 two cited, May 18 five-plus cited. The trend is real, the durability framework appears met, but the magnitude could regress on the May 25 rotation.

We'll know more then. I'd rather publish this piece now and update it in two weeks than wait for total certainty and miss the news window on the Google guide. That's a tradeoff I'm making consciously.

What's changing in the playbook — and what isn't

To be clear about what we're updating in Part 1 after this week's data:

The framing that you "must add" FAQPage JSON-LD is being softened. Keep it if you have it — it earns SERP rich-result eligibility, which still drives standard search clicks. Don't expect it to drive AI Overview citation by itself.

The framing that dated freshness drives AI Overview citation is being explicitly retracted. Dated freshness drives standard SERP signals. AI Overview doesn't separately reward it.

The seven-element format pattern is becoming an eight-element pattern, with the new eighth element being "AI Overview synthesis is decoupled from crawl freshness."

Here's what isn't changing:

• The answer-first prose pattern. The discipline of writing a 134-to-180-word answer at the top of every pillar. This is still where the citation work gets done.
• The semantic HTML structure. H2/H3 sectioning is load-bearing.
• The named-entity authority sourcing. Institutional citations make the pillar synthesizable.
• The internal-link mesh.
• The 30-day field-report cadence. We'll keep publishing data, because the data is what makes the operating system real instead of vibes.

The shorter version of all of this is the line I want to leave with you:

The discipline isn't in the markup. It's in what you put on the page before you mark it up.

---

Frequently asked questions

Should I rip out my FAQPage JSON-LD?

No. Google's guide says structured data is "not required for generative AI search" — not that it's harmful. The FAQPage JSON-LD still earns SERP rich-result eligibility, which still matters for standard clicks. What changed is the expectation that the schema is doing the AI Overview citation work. It isn't. Keep it for the SERP eligibility; stop banking on it for AI Overview.

Does your 30-day field report still hold?

Mostly, yes. Five of the seven format-pattern elements — answer-first prose, stat density, internal-link mesh, semantic sectioning, named-entity authority — hold up under this week's data. Two elements (FAQPage JSON-LD as load-bearing, dated freshness as a separate signal) are being explicitly softened or retracted. A new eighth element — that AI Overview citation is decoupled from crawl freshness — is being added.

What's the new format pattern, in one paragraph?

Lead every pillar with an answer-first prose block of 134-to-180 words at the top. Maintain stat density of about one named-source citation per 200-300 words. Use semantic HTML with clear H2/H3 sectioning. Build an internal-link mesh of four to nine links to related pillars. Cite institutional sources — journals, regulators, research firms — for authority. Keep the FAQPage JSON-LD for SERP rich-result eligibility. Skip the content-chunking, the AI-specific rewriting, the inauthentic-mention chasing.

Why is your data telling a different story than the consultancies recommending heavier schema?

We have production data on indexed pillars across a 30-day window. Most consultancy recommendations are based on testing on hypothetical pages, or correlating SERP outcomes against schema presence without a control group. The pillar earning peak citations this week is itself most of a control group — it has the FAQPage JSON-LD attached, but Google's rich-result detector isn't currently surfacing any FAQ markup for it, and it's earning more citations than ever. The schema isn't where the action is.

What does this mean for the Claude Cowork pillars?

Same answer-first pattern. Same semantic sectioning. Same named-entity authority. The FAQPage JSON-LD is optional. The bigger lever is leaning into the practitioner-voice content gap that Anthropic webinars and vendor pages like Spellbook and Harvey can't fill. The Cluster 7 pillars we shipped May 15 — the law-firm piece and the M&A due-diligence piece — are the test bed. The May 25 rotation will tell us whether they're earning citations the same way this pillar did.

What changes about how we work, day-to-day?

Less time on schema-pattern fiddling. More time on prose discipline and brand-authority signals. The "publish and let the recompute find it" model implies we should be writing more pillars and refreshing less. A working-session week that produces one new pillar at the discipline level we maintained on this pillar is probably better leverage than three weeks of schema A/B testing on the same pillar.

---

Sources

1. Google Search Central — Optimizing your website for generative AI features on Google Search (2026-05-15)
2. Google Search Central Blog — A new resource for optimizing for generative AI in Google Search (2026-05-15)
3. Search Engine Journal — Google's New AI Search Guide Calls AEO And GEO 'Still SEO' by Matt Southern (2026-05-15)
4. The Keyword (Google blog) — AI Mode in Google Search and AI Overviews get Gemini upgrades (2026-05-14)
5. web.dev — Agent-friendly website best practices
6. UCP — Universal Commerce Protocol (ucp.dev)
7. Our own AI Overview citation: 30-day field report (Part 1)
8. Our own data: DataForSEO serp_live captures from 2026-05-11 through 2026-05-19, archived in our internal aio-rotation/what-is-creative-technology-agency.md