How three of our pillars broke into AI Overview citation in 30 days (a field report)

In thirty days, three of our ten target queries for Answer Engine Optimization (AEO) flipped from zero citations to AI Overview citation. Here's exactly what shipped, what indexed, what got cited, and the format pattern we extracted. The headline finding: pillar-grade pieces with a self-contained 134 to 167 word answer block at the top, structured statistic density, FAQPage JSON-LD (a Google-recognized structured-data format for FAQ sections), and dated freshness all earn citation pickup within 1 to 3 weeks once they are indexed. The format mattered more than the SEO score.

A second finding: ranking and citation are different problems with different solutions. We rank in the top 10 on a brand query while being conflated with an unrelated open-source project. We don't rank top 10 on the creative-technology-agency query yet, but we get cited 8+ times in the AI Overview for the same query. The data below is the cleanest practitioner field report of an AEO citation breakthrough we've been able to find in public. Most of what's published is framework-grade speculation, not program-level data.

Why publish this

Most of what's published about getting cited by AI Overviews falls into one of two categories. The first is software-vendor explainers from tools like Profound, AthenaHQ, or SE Ranking, selling the tool that claims to optimize for citation. The second is framework-grade SEO-publication essays about how to think about AEO. Neither category publishes program-level field data. Vendors don't have agencies running their own publishing programs to draw from. SEO publications can sketch the framework, but they don't run the experiments. Practitioner agencies that do run the experiments tend to treat the data as proprietary.

We're publishing because the report we wanted to read 30 days ago, but couldn't find, is the same report we can now write. And because Google's AI Overview outbound-link update on May 6, 2026 explicitly named "Expert Advice" as a new citation surface that the algorithm now expands. First-hand practitioner data is exactly that surface. The report writes itself; we'd be silly not to.

The 30-day timeline

We started the publishing cadence on April 13, 2026. The first three weeks were technical foundation: sitemap, canonical, indexing. The publishing cadence proper started April 15. Here's what shipped, when, and what happened.

Date	Piece	Status as of 2026-05-08
2026-04-15	"What a creative technology agency actually does"	Indexed; Q1 cited 8+ times in AIO today
2026-04-15	"AI agency vs traditional agency: why the comparison is wrong"	Indexed; Q3 cited 3+ times in AIO today (Compounding Value paragraph single-sourced)
2026-04-17	"AI Automation ROI: What to Realistically Expect in 2026"	Indexed; AIO multi-cite (5+) including entire "Where to Focus" list lifted verbatim
2026-04-20	"How we build: the Automaton stack"	Indexed; 1,077 impressions in 28 days at position 8.90, with a brand-disambiguation problem (the AI Overview is conflating us with a different "Automaton")
2026-04-20	"5 real challenges small businesses hit with automation (2026)"	Indexed (first to break through, via legacy 301 equity); content-decay flag pos 2.4 → 5.67
2026-04-24	"Five-layer framework AEO patch"	Indexed; Q7 cited 5+ times in AIO (closing failure-mode paragraph single-sourced)
2026-04-27	"AI for accountants: what actually works in 2026"	Indexed; ranking pos 6.41 with 175 impressions over 28 days
2026-04-29	"AI receptionist for small business: build-vs-buy"	Indexed
2026-05-06	"Claude Cowork vs Claude Code"	Not yet indexed (URL unknown to Google as of 2026-05-08; manual Request Indexing pending)

That's nine pillar-grade pieces shipped in 23 days, plus five older blog posts that pre-date the pillar program. Three of the new pieces are actively cited in AI Overviews today. One older piece is also cited (the five-layer-framework piece, which got an AEO patch on April 24 and broke into AIO citation by May 7). The supplementary ai automation roi what to expect 2026 query, which sits adjacent to one of our pillars, has gone from zero cites plus a Featured Snippet last week to five-plus AIO cites this week, with the entire "Where to Focus" list lifted verbatim from our piece.

What the cited pieces have in common

We pulled the rendered structure of each cited piece and the SERP-side AIO body text to extract the format pattern. Here's what the four AIO-cited pieces share.

A self-contained BLUF (Bottom Line Up Front) answer block of 134 to 180 words near the top. Each answer block opens with a one-sentence definitional statement of the topic, follows with three to five fact-dense supporting points, and closes with a direction-setting sentence about what the rest of the piece will cover. The answer block reads as a self-contained mini-article. An algorithm doing a passage-level pull can lift it cleanly without needing context from the surrounding piece.

FAQPage structured data with four to seven fan-out questions. Each piece has an FAQ section at the bottom marked up in JSON-LD, where the question text matches the kind of related queries the AIO algorithm is expanding on the parent search results page. The answers are two to four sentences each, fact-dense, attribution-anchored where statistics are involved.

Statistic density with explicit attribution. Specific 2026 statistics (McKinsey on agent-pilot failure rates, Gartner on data quality, MIT on enterprise AI ROI failure, BCG on agency-AI adoption, AICPA on accountant AI adoption) appear at a rate of roughly one statistic per 200 to 300 words of body text, each with its source named explicitly. The AI Overview algorithm preferentially lifts attribution-anchored claims, because they're easier to evaluate as authoritative.

Dated freshness. Each cited piece has a visible publication date and, where applicable, a last-updated date. Multiple pieces in our published-and-cited set have had dates bumped within the last 30 days as part of freshness operations. The pattern that 95% of ChatGPT citations come from content updated in the last 10 months, which is commonly published in the AEO literature, is consistent with what we're observing on the Google AIO side too.

Internal-link mesh. Each cited piece links to four to nine other Automaton pieces. The internal-link mesh reinforces topical authority. It's the kind of "this site is a coherent body of work on this topic" signal that AI Overviews seem to prefer to lift from.

Hero image with brand-prefix B&W aesthetic. Less directly tied to citation, but the consistent visual treatment makes the brand recognizable, which contributes to the entity-recognition layer that's increasingly important for AI engine source selection.

Behind the format: how the program actually runs

The format pattern above is the what. The cadence and the human-versus-Claude division of labor is the how. We're being explicit about this because it's the part that vendor explainers and SEO publication essays both leave out, and it's the part most practitioners actually need.

This program runs inside Claude Cowork. Cowork is Anthropic's product that lets non-developers run agentic work as scheduled tasks, with plugins that give the agent specific capabilities. Our setup uses three Anthropic-built and custom-built capabilities. DataForSEO via a 12-tool MCP (Model Context Protocol, the standard Anthropic uses for plugin integration) covers keyword research, search results page analysis, competitor research, and on-page audits. Google Search Console via a 9-tool MCP covers indexing inspection, striking-distance keyword detection, content-decay scanning, cannibalization checks, and click-through rate analysis. The Automaton Content MCP covers reading and writing pages and blog posts on the live site, plus image generation and voice-clone tooling.

The agent runs inside Cowork on a published cadence. Daily, the agent does a five-minute health check Monday through Friday: scan algorithm news, run a rotating two-query AI-citation check, diff the live site against the inventory, push any auto-tier fixes that are queued, post a short Slack summary. Working sessions happen three to four times per week, themed: vertical-pillar work on Mondays, build-versus-buy and Cowork-cluster work on Wednesdays, infrastructure work on Fridays. Each working session is constrained to ship one deliverable. Bi-weekly, the agent runs the full strategic review you're reading the output of right now: ten-query citation retest, full GSC analytics, brief generation, strategic plan refresh.

The human input layer is Joseph. Joseph sets cluster strategy, decides which verticals are co-equal #1 priorities, runs the Strategic Assumption Registry that says what bets the program is making, gives voice red-lines on every draft (the first Cowork pillar we shipped went through a v1-to-v2 rewrite because the v1 voice was too technical for the dual-persona target audience), and makes the publish go-or-no-go call on every piece that crosses the 30%-structural-change threshold. Joseph also runs the meta-audit: a separate scheduled review where the agent's own work product is examined to ask "is this still the right work?"

The agent's role is research, analysis, drafting, and execution of the cadence. The agent does not make strategic decisions. The agent doesn't decide which queries to target, which verticals to invest in, when to graduate a piece from draft to publish, or how to sound. The agent runs the system Joseph designed, surfaces what the data says, drafts what the strategy says to draft, and waits at the red-line. When Joseph approves auto-tier fixes (small structural changes, internal-link sweeps, freshness date bumps), the agent ships them directly via the Content MCP and triggers a Vercel deploy. When the work is approval-tier (new pieces, major edits, schema changes, anything over 30% structural), the agent queues a draft and a preview, and waits for Joseph.

In practice, a typical week looks like this. Daily five-minute checks run autonomously; Joseph reads the Slack summary on his phone in two minutes. Mondays, Wednesdays, Fridays, the agent ships a brief or a draft to the Slack channel. Joseph reviews drafts in 20 to 40 minutes per piece. Joseph's actual time investment is roughly 2 to 4 hours per week on the program, almost all of it spent on strategic decisions and voice review. The agent's running time is roughly 10 to 15 hours per week of compute, doing the research, drafting, and analysis that would otherwise consume 20 to 30 hours of an SEO operator's time. The leverage ratio is high because the work that compounds, the strategic decisions and the voice integrity, stays human.

The single most important point: the agent reaches its limit at the 30%-structural-change boundary. It cannot ship a new pillar autonomously. It can refresh a date, fix a typo, add an inline link, or push a queued auto-tier change without human review. It cannot decide that a piece deserves to exist, or that a piece's voice is right, or that a piece is ready to publish. Those decisions stay with Joseph by design. The result is a system that compounds steadily without compounding into the wrong work.

The case study at /work/seo-aeo-engine is the engineering view of the same system: which Model Context Protocols expose which tools, how the recurring task is configured, how the Slack hand-off works, what the data flow looks like end-to-end. This blog post is the field-report view of what the system produced in its first 30 days. They're complementary; read them together.

What the uncited pieces are missing

Six of the ten formal AEO target queries are still 0/10 cited (Q2, Q4, Q5, Q6, Q8, Q9, Q10). The gap analysis breaks them into three sub-categories.

Hard-incumbent queries (Q2, Q4, Q6). The AIO citation surface on these queries is locked by institutional and platform giants: BCG, AWS, IBM, Microsoft, McKinsey, MIT Sloan, the Small Business Administration, the U.S. Chamber of Commerce, Wix, Squarespace. At our domain authority, which is effectively zero (DataForSEO's ranked_keywords API still returns zero for our domain, meaning we're below the data-set threshold), we can't realistically crack these. We accept the gap and don't waste resources on retrofits we can't win.

Format-mismatch queries (Q5, Q8). Our existing pieces on these topics are indexed and structurally eligible. The Q3 query cites our ai-agency-vs-traditional-agency piece three times, which means the algorithm has read our content as authoritative. But the pieces don't slot into the answer rhythm the AIO is harvesting on Q5 and Q8 specifically. Q5 wants a definitional answer block; ours is thesis-shaped. Q8 wants a yes-or-no-or-nuance framework; ours is an inversion-frame ("the comparison is wrong"). These are retrofittable, and we shipped a multi-piece retrofit brief this week to handle them.

Ideal-customer-profile mismatch queries (Q9). "What is a creative technologist" is a career-explainer query in 2026. The search results page is dominated by NYU FRL, Cella, freeCodeCamp, ZipRecruiter, Wikipedia. Our piece is positioned for a different reader, the small-business owner considering an agency, than the search results page actually serves, which is the early-career professional considering a career path. The 10-query target list might need a buyer-shaped query swap rather than a retrofit.

The single piece that proves the format works

The ai-automation-roi-what-to-expect piece is the cleanest format-validation case in the program. Last week (May 1, 2026), the piece owned a Featured Snippet on ai automation roi what to expect 2026. But the AI Overview was citing Ringly, WNDYR, SEM Nexus, and MIT-95% statistics from sources we outranked. We had the ranking; we didn't have the citation pickup. That gap was the single most important strategic question of the May 1 weekly report.

This week, with no further structural editing (the piece has not been touched since publication), the AI Overview has reorganized to cite Automaton five-plus times across four body sections. The opening paragraph cites our 300%-to-330% ROI claim (paired with thelead.io). The "Realistic ROI & Payback Expectations" section cites our Small Business Advantage 280%-to-520% range. The "Hybrid Human-AI Workflows" section cites our piece. And the entire "Where to Focus for Maximum ROI in 2026" list (Customer Service, Back-Office, Sales & Marketing, Content Production) is lifted verbatim from our piece.

What changed? Indexing matured (the piece had been indexed for about three weeks); the algorithm had time to evaluate the content's authority; the format pattern that we'd built in did the work. No retrofit was needed. That's the data point that makes the format pattern claim defensible. When the format is right and the indexing is mature, the algorithm picks the content up.

The counter-finding: ranking is not citation

Our automaton-stack pillar is indexed, ranks position 8.90 for a brand-adjacent query (automaton ai agent framework), and pulls 950 impressions per 28-day window with zero clicks. The AI Overview for that query describes a different "Automaton" (an open-source crypto-native AI agent framework hosted on GitHub, sourced from aiagentstore.ai). The AIO is conflating two unrelated entities.

The fix isn't a content edit. The fix is entity disambiguation at the structured-data layer: the home page Organization JSON-LD that we shipped as a brief on May 1 (and that's still pending Joseph's repository push). Once the structured data is live and Google re-indexes the home page with explicit entity assertion (founder, address, sameAs links to social profiles, knowsAbout, makesOffer), we expect the disambiguation to resolve over a 14-to-30 day window.

The lesson: ranking, citation, and entity recognition are three different problems with three different fixes. Ranking is content plus backlinks plus freshness. Citation is the format pattern we've described. Entity recognition is structured data plus brand consistency plus, eventually, a Wikidata entry and a Google Knowledge Graph entity. A piece can succeed at any one and fail at another. Plan for all three explicitly.

The 5-step playbook for any client at our DR (essentially zero)

For any practitioner reading this who's running a publishing program at a small DR (Domain Rating), here's the playbook synthesized from what we've watched work.

1. Pillar-grade word count: 3,000 to 4,500 words per pillar. Not because length itself is rewarded, but because the depth required to earn citation pickup compresses into roughly that range. Below 2,500 words, the piece typically lacks the statistic density and fan-out coverage to be citation-eligible.
2. Self-contained BLUF answer block: 134 to 180 words near the top, structured as one definitional sentence plus three to five fact-dense supporting points plus one direction-setting sentence. Reads as a self-contained mini-article that an algorithm can lift cleanly.
3. FAQPage JSON-LD with fan-out questions: four to seven questions, structured to match the kind of expansion queries that AI Overviews are running on the parent search results page. Answers two to four sentences each, fact-dense, attribution-anchored.
4. Statistic density with explicit attribution: one stat per 200 to 300 words, with explicit source naming. The algorithm preferentially lifts attribution-anchored content because it's easier to evaluate as authoritative.
5. Dated freshness plus internal-link mesh: visible publication date, visible last-updated date, four to nine internal links to related pieces.

That's the format. None of it is hidden. None of it requires expensive tooling. The discipline is the implementation, not the technique.

What we don't know yet

Honest section. AI Overview composition is intermittent. The same query can produce a different cited set on different days. One week's pattern may not hold next week. Google's 2026-05-06 update added five new outbound-link surfaces to AIOs ("Explore new angles," "Expert Advice," and three more), and how those surfaces will redistribute citation share is still being learned. Anthropic's Claude (the underlying model) is also competing as an AI engine, and the citation surfaces there don't yet correlate cleanly with Google AIO patterns.

The May 15, 2026 Chrome DevTools-baseline session we have scheduled will close one of our remaining measurement gaps: the rendered-page-truth check that our format pattern claim partially rests on. If the data surfaces something the search-results-side data missed, we'll publish a correction. The data is the data; we report what we have.

What this means for SMB owners considering AI agencies

If you're not a practitioner, if you're a small-business owner reading this, trying to figure out whether AI is going to actually work for your business, the bottom line is simple. The agencies that publish program-level data, run experiments, and document what worked versus what didn't are operating in a fundamentally different way than the agencies that publish marketing-grade thought leadership without measurement. The first kind compounds. The second kind doesn't. Pick the agencies that can tell you what they ran and what happened.

If you want to see the system that produced this data, the SEO/AEO engine case study is the engineering view of the same program: which Model Context Protocols expose which tools, how the cadence runs, the human-versus-Claude division of labor, and the build-vs-buy framework we apply to every category. This blog post is the field-report view of what happens when we run it. They're complementary; read them together.

We're available for a conversation if your firm is at the point of evaluating whether to invest in this kind of program for your own publishing surface.

FAQ

How long does it take for an AI Overview to cite a new piece of content?

Based on our 30-day program, citation pickup typically arrives one to three weeks after the piece is indexed by Google, assuming the content matches the format pattern (BLUF answer block, FAQPage JSON-LD, statistic density, dated freshness). The bottleneck is usually indexing. Google has to crawl, evaluate, and decide the content is authoritative before AI Overview source selection considers it. Once indexed, the algorithm pulls candidate citations on a roughly weekly cadence based on what we've observed.

Do you need backlinks to get cited by AI Overviews?

Not in the same way you need them for organic ranking. Our domain returns zero ranked keywords in DataForSEO's data set (meaning we're below the data-set threshold for measurable backlink-driven authority), and three of our pillars are still cited multiple times in their respective AI Overviews. The format pattern and the practitioner-voice gap on a given query matter more for citation pickup than backlink authority does, at least at the small-DR end of the spectrum.

Does FAQPage schema help with AI Overview citation?

Yes, in our observation. Every cited piece in our program has FAQPage JSON-LD with four to seven fan-out questions matching the expansion queries AI Overviews run on the parent search results page. We can't run a controlled test (no piece without FAQPage schema to compare against), but the pattern is consistent enough that we treat FAQPage schema as a required element of the format pattern.

What kind of content gets cited most often in AI Overviews?

Based on what we've watched: practitioner field reports on niche topics where institutional voices haven't locked in the citation surface; pillar-grade pieces with self-contained BLUF answer blocks; content with explicit statistic attribution; content updated within the last 10 months. The opposite (thin content, anonymous voice, no schema, no dated freshness) almost never gets cited regardless of organic ranking.

How is Google's AI Overview citation different from a Featured Snippet?

A Featured Snippet is a single-source answer pulled from one URL into a position-zero answer box. An AI Overview citation is a multi-source citation pulled from multiple URLs into a structured answer surface. The Featured Snippet algorithm rewards stable, well-formatted answer blocks that match the query intent precisely. The AI Overview algorithm rewards passages that the model considers authoritative and that slot well into the structured-answer composition. We've watched a Featured Snippet on one query convert into AI Overview multi-cite over a seven-day window. Net citation surface is stronger after the conversion.

Can a small agency with no domain authority compete in AI Overview citation?

Yes, for queries where the citation surface has a practitioner-voice gap. Hard-incumbent queries, where institutional and platform giants have locked the citation surface, remain uncrackable at small DR. But mid-tier and long-tail queries where the search results page is split between vendor-explainers and journalist content frequently have an empty practitioner-voice slot. That's where small-DR practitioner agencies can earn citation pickup with the format pattern. We documented three such pickups in 30 days, and we publish the data so practitioners reading this don't have to discover it themselves.