Found in AI Ep. 52: A Third AEO, Google’s AI Contribution Pilot, and the NYT Accuracy Study

TL/DR: Four major AI search stories landed between April 13 and April 15, 2026: Google Cloud’s Addy Osmani introduced a new framework called Agentic Engine Optimization (AEO), Google is piloting an AI Contribution Report inside Search Console, a NYT-commissioned study put hard numbers on AI Overview accuracy and an ungrounding problem that got worse with Gemini 3, and Google Ask Maps is shifting local search from listings to recommendations. All four stories point to the same shift: AI search infrastructure is hardening.

This week’s edition of Found in AI recaps four stories that, taken together, signal a real shift in AI search. The infrastructure around generative engines is getting measurable, optimizable, and auditable in ways it wasn’t a year ago. Here’s what happened and what marketers should actually do about it.

Story 1: Agentic Engine Optimization — A Third AEO Enters the Chat

Quick explanation: Addy Osmani, a director of engineering at Google Cloud AI, published a new framework called Agentic Engine Optimization (AEO) on April 15, 2026. It focuses on optimizing content for AI agents—not search engines, not answer engines—AI agents that fetch, parse, and act on content autonomously.

This is now the third “AEO” in circulation. Marketers already use the acronym for:

• Answer Engine Optimization: optimizing for AI-generated answers
• Artificial Intelligence Optimization: often used interchangeably with GEO
• Agentic Engine Optimization: Osmani’s new framework, focused on AI agents

The naming situation is a mess. But the framework itself is useful.

What Osmani is saying

AI agents don’t consume content the way humans do. They don’t scroll, click, or engage with a UI. They fetch, parse, and act—all in a single request. That makes most traditional engagement metrics irrelevant for agent-driven visits.

More importantly, every AI agent operates inside a finite context window. If a page exceeds that window, one of three things happens:

1. The agent truncates the content
2. The agent skips the page entirely
3. The agent hallucinates an answer

For brands, option three is the most dangerous. An agent might confidently misrepresent what a business actually offers.

What to optimize for

Osmani’s recommendations for Agentic Engine Optimization:

• Put the answer early. Within the first ~500 tokens, ideally.
• Keep pages compact and focused. One page, one job.
• Skip long preambles. Agents have “limited patience” for narrative intros.
• Serve clean Markdown alongside HTML. Markdown strips navigation and script noise, making content cheaper to parse.
• Consider structured discovery files like llms.txt, skill.md, and AGENTS.md as machine-readable entry points.

Important caveat

This type of AEO is not about Google Search rankings. Google’s John Mueller has publicly stated that Google does not recommend separate Markdown pages and does not use llms.txt as a ranking signal. Agentic Engine Optimization is about making content usable inside agent workflows—Claude, ChatGPT agents, and other tools that fetch pages to complete tasks.

What to actually do about this

Audit the highest-value commercial pages on your website—pricing, comparison pages, service pages. On those specific pages, confirm the core answer appears within the first couple hundred words. Tighten anything that buries the point under long mission statements or hero content.

Don’t rush to build llms.txt or AGENTS.md files across your entire site. The standards are still moving. Structural clarity on your existing content matters more right now.

Story 2: Google’s AI Contribution Report in Search Console Is In Pilot

Quick explanation: On April 13, 2026, Barry Schwartz of Search Engine Roundtable spotted references to an “AI contribution pilot” inside Google’s own support documentation. This appears to be a new Google Search Console report that tracks how content performs inside AI-generated search experiences.

There are no public screenshots yet. No official announcement. What exists is documentation referencing the pilot, plus a prior hint from Google’s John Mueller in early February suggesting something like this was coming.

Why this matters

Microsoft’s Bing Webmaster Tools launched an AI Performance dashboard earlier this year that tracks:

• Total citations in AI answers
• Which URLs are being cited
• Grounding queries
• Page-level citation activity
• Citation trends over time

If Google’s AI Contribution Report follows a similar model, publishers would gain the ability to:

1. Separate AI Overview visibility from traditional blue-link visibility. Something currently difficult inside the standard Search Console Performance report.
2. Identify the pages most used as source material for AI-generated answers. A direct authority signal.
3. Build a feedback loop. Up to now, AI search optimization has been largely a “publish, prompt-test, hope” exercise. A dashboard turns it into something measurable.

What this signals

The fact that Google is testing this at all is more significant than the feature details. Google was slow to give publishers anything around AI-generated search, moving from “trust us, normal SEO works for AI Overviews” to “here are some controls” to — now — “here’s how you can measure it.” That progression matters.

Story 3: The NYT Study on AI Overviews Accuracy

Quick explanation: The New York Times commissioned AI startup Oumi to run ~4,300 Google searches through the SimpleQA benchmark. AI Overviews powered by Gemini 2 were 85% accurate. After Google upgraded to Gemini 3, accuracy climbed to 91%. More than half of “accurate” responses were ungrounded—meaning the answer was right, but the cited sources didn’t fully support it.

The headlines focused on the accuracy number. The more important finding is buried underneath it.

The scale problem

Google handles around 5 trillion searches per year. Even a 9% error rate, at that scale, produces tens of millions of incorrect answers every hour—hundreds of thousands every minute.

The ungrounding problem

More than half of the AI Overview answers that were technically accurate cited sources that didn’t fully support the information presented. The answer was right. The attribution was wrong. And this ungrounding issue appears to have gotten worse with the upgrade from Gemini 2 to Gemini 3, not better.

The manipulation problem

Lily Ray of Amsive ran a stress test. She published an AI-generated article titled something like “The Top 10 Experts at Knowing What Google Really Wants in 2026.” Within a day, Google was citing that article inside AI Overviews. Her point: the manipulation problem is not confined to edge-case queries. High-intent commercial queries—the ones buyers actually use before converting—are among the most heavily gamed.

Google pushed back, calling the study flawed and noting that SimpleQA is not representative of real search behavior.

What this means for the FSA Framework

Inside the FSA Framework (Freshness, Structure, Authority), the Authority pillar is about being referenced consistently and credibly across verifiable sources—not just being mentioned everywhere. The ungrounding findings reinforce why:

• Authority is entity density, not volume. Consistent mentions across trusted third-party sources outweigh any single self-declared claim.
• Grounding matters for retrieval. Structure content so claims can be traced to sources. Cite data. Reference frameworks. Name tools and standards.
• Short-lived manipulation gets filtered. The brands that compound authority over time are the ones still cited when retrieval systems tighten.

Story 4: Google Ask Maps Is Moving From Listings to Recommendations

Quick explanation: Search Engine Land published a hands-on test of Google Ask Maps on April 14, 2026, focused on local service queries (plumbers, electricians, HVAC companies). The finding: as prompts become more nuanced, Ask Maps stops listing businesses and starts recommending them based on qualities like responsiveness, specialization, and honesty.

The pattern

At the simple end of the query spectrum—”plumber near me”—Ask Maps behaves like traditional Google Maps. A ranked list of nearby businesses.

As queries get more conversational—adding uncertainty, trust signals, decision-making language—Ask Maps narrows the field. It interprets user intent and frames businesses around qualitative attributes:

• Responsiveness
• Specialization
• Honesty
• Repair-first thinking

For complex prompts, Ask Maps sometimes offers guidance before recommending any businesses at all.

What to do about it

Local search is now going through the same shift the rest of search already went through—moving from retrieval to recommendation. The business ranked #1 for “plumber” is not necessarily the business recommended when a user asks for “an honest plumber who won’t upsell me.”

Four practical actions:

1. Reviews matter more, not less. The qualitative language inside reviews shapes how the model describes your business.
2. Google Business Profile descriptions need specificity. “Full-service HVAC” is weaker than “Specialists in older homes, focused on repair before replacement.”
3. Audit what Ask Maps says about you. Ask it a few real-world prompts your customers might use. If the framing is off, that’s a reviews-and-content problem to work on.
4. Update agency reporting. Map pack position isn’t enough. Include a regular Ask Maps check on how AI describes local clients in natural-language prompts.

The Through-Line: AI Search Infrastructure Is Hardening

Taken together, these four stories point in the same direction:

• A Google engineering director is publishing frameworks for how agents consume content.
• Google is building native AI visibility reporting inside Search Console.
• Researchers are quantifying AI accuracy and the ungrounding problem.
• Local AI search is moving from retrieval to recommendation.

Every one of these is infrastructure. And when infrastructure solidifies, behavior follows.

AI search visibility is now measurable, optimizable, and auditable in ways it wasn’t a year ago. The brands that treat this as an opportunity will pull away from the ones treating it as noise.

If we haven’t met yet….

Hi, I’m Cassie, a fractional content strategist for startups and enterprise brands who focuses on AI search optimization.

I build content programs that connect strategy with execution: clear positioning, systems, and content that actually drives revenue. If you’re ready to stop guessing and start growing, here’s how we can work together.