We pointed an autonomous agent at our own SEO for 90 days. Here's the citation data.

For about 90 days, an autonomous agent has run our own SEO and AEO program end to end — it pulls Search Console and analytics data, audits AI Overview citations, drafts and ships content, and reports to us every morning. The single most useful thing it produced is not a ranking. It's a metric we now call the citation ledger: not "do we rank," but "when an AI Overview appears on a query we care about, does it cite us — and is that trending up or down." Over the window, organic sessions went from a near-flat line to roughly 179 per 28 days, AI assistants became a measurable traffic channel for the first time (about 36 sessions per 28 days from ChatGPT, Perplexity, Gemini and Google's AI surfaces), and our share of cited target queries moved off zero. The honest headline, though, is not the growth. It's that ranking and being cited have become two different games — and most of what we learned was about the second one.

This is a field report, not a how-to. If you want the implementer's playbook, read how to rank in AI Overviews and our guide to generative engine optimization. This piece is the other thing: what running the engine on ourselves, daily, actually taught us — including the parts that didn't work.

What "an autonomous agent runs our SEO" actually means

It's easy to say "we automated our SEO" and mean a few scheduled reports. We mean something more literal. The program runs as a scheduled task that fires every day. It checks that its data connections are live, then branches by day of week: a light health check most days, a working session twice a week, and a full deep dive on Fridays. On a given run it queries Search Console for trends, decay, striking-distance keywords, and cannibalization; runs live searches to read the actual AI Overview on our target queries and record who got cited; audits our internal link graph for orphaned pages; drafts content; ships the safe, small changes itself; and queues the bigger ones for a human. Then it writes a report and posts it to Slack. We describe the architecture behind it in the SEO/AEO engine case study, and the broader stack it runs on in the Automaton stack.

The important design choice: the agent has real authority over small, reversible changes — internal links, answer-block tweaks, freshness updates, re-index requests — and zero authority over new pages or whole-section rewrites, which always wait for a human. That boundary is the whole game. It's what lets the thing run unattended without becoming a liability. We didn't arrive at it cleanly; we tightened it after watching what the agent was good at (relentless, consistent, data-driven small moves) and what it shouldn't touch (judgment calls that change the argument of a page).

The metric that actually matters: the citation ledger

Here is the reframe that changed how we run the program. In the classic model you track rankings and clicks. In the AI-answer era those two numbers quietly stopped telling the truth, because an AI Overview can sit above your number-one result and answer the question so completely that nobody clicks. You can hold rank and lose traffic. You can lose rank and still get cited inside the Overview, which is its own kind of win. Rankings and clicks measure a world that's being eaten by the answer box.

So we built a different instrument. For every query we care about, each audit records two things: was an AI Overview present at all, and — if it was — did it cite us. Track those over a trailing window and you get the citation ledger: an AIO-present rate (how often the answer box even shows up on this query) and a cited-when-present rate (how often we're in it when it does). That second number is the one that predicts whether AI assistants send you traffic, and it moves independently of your blue-link rank. We've watched a page sit at organic number one with the Overview absent one week and present-but-citing-someone-else the next. The ledger captures that; a rank tracker doesn't.

If you take one idea from this piece, take that one. Stop asking "do we rank for this." Start asking "when the machine answers this question, is our name in the answer, and is that getting better or worse." For more on why this is a different discipline from classic search, see GEO vs SEO: what actually changed.

What 90 days actually produced

The numbers, honestly. We started the program with organic traffic that rounded to nothing — a dozen clicks across a month, an impression count in the hundreds. Across the window the trend reversed: organic sessions reached roughly 179 per 28 days, with impressions in the thousands per day at the peak. More interesting than the raw growth is what showed up in the analytics for the first time: a named "AI Assistant" traffic channel. Around 36 sessions per 28 days now arrive from ChatGPT, Perplexity, Gemini, and Google's AI surfaces — traffic that simply wasn't attributable before, and that exists because of citations, not classic rankings. The same week that channel became legible, our conversion events started firing for the first time too. None of these are vanity-tier numbers for a giant brand; they're the honest early-stage numbers of a small site, which is exactly why we'll show them. The shape is the point.

On the ledger itself: we went from zero cited target queries to a handful held over multiple weeks — won, lost, and in some cases won back. That last category matters more than it sounds, because it taught us the most.

Five things running the engine taught us

1. The AI Overview will adopt your exact framing — and cite someone else

This is the most disorienting thing we've watched, and it happened again the week we published this piece. On one of our highest-traffic queries, the AI Overview now explains the topic using a distinction we wrote — close to our exact phrasing — and then cites two other sites for it. The idea propagated; the attribution didn't. The lesson is brutal and clarifying: being the source of an idea does not make you the cited source. Citation goes to the page that states the claim most cleanly, most recently, and in the most liftable format on the day the Overview recomputes — not to whoever said it first. We responded the only way that works: we made our own page state the claim more cleanly and more currently than the pages getting the credit. You don't win this by being right earlier. You win it by being the most liftable on the day.

2. Ranking number one and getting cited are decoupled

We have pages that rank at the top and aren't cited, and pages cited inside the Overview that don't rank in the top ten beneath it. Google assembles the answer box from whoever has the cleanest passage for each sub-claim, and that selection is only loosely correlated with classic ranking. Practically, this means two separate workstreams: a ranking workstream (the old discipline) and a citation workstream (answer blocks, FAQ structure, first-hand statements, freshness). Optimizing one does not automatically deliver the other. Most teams are still only running the first.

3. A platform change can rewrite your citations overnight

In early May, Google shipped an update that started pulling forum posts and first-person "expert advice" into AI Overviews far more aggressively. Our citation ledger registered it as a step-change: on several queries, single-source practitioner citations (ours) got displaced by Reddit threads and discussion blocks the same week. We didn't lose because our content got worse. The selection criteria moved. The takeaway is that in the AI-answer era, your citations sit on top of a platform that can re-weight what it rewards without warning — so the monitoring has to be continuous, not quarterly. A ledger you read every day catches this; an audit you run every quarter explains it to you three months late.

4. On a small site, getting crawled is the real bottleneck — not "quality"

We published a batch of strong pages on a single day and watched most of them sit unindexed for over a week — "discovered, not indexed," in Search Console's phrasing. Nothing was wrong with the pages. We had simply shipped a wave of near-orphaned URLs into a small site with a modest crawl budget, and Google didn't prioritize discovering them. The fixes that actually worked were unglamorous: internal links from pages Google already crawls often, and a manual re-index request for the genuinely stuck ones. We now treat internal linking as a crawl-discovery system, not a relevance nicety, and we never ship a wave of unlinked pages at once. If you're a small site wondering why your good content isn't ranking, check whether it's even indexed before you touch the content.

5. A single-topic competitor will out-rank you on definitional queries — so don't fight there

On the cluster where we get the most traffic, the page that consistently beats us on the definitional and pricing queries is an exact-match-domain microsite whose entire domain is about that one topic. We're an agency where that topic is one of several. On pure topical concentration, the single-topic domain wins the definitional fight, and no amount of refresh fully changes that. So we stopped fighting there. We compete where the agency actually has an edge: accuracy and currency (their pages drift stale; ours don't), genuine first-hand experience (the "expert advice" the platform now rewards, which an anonymous guide site can't fake), and the judgment queries — "is it worth it," "when not to use it," real cost and ROI — that convert and that descriptive microsites don't write. You can read that thinking in practice in our Claude Cowork pricing breakdown and the field guide for law firms. Pick the fight you can win and that pays.

What we'd tell anyone pointing an agent at their own search

Three things. First, instrument the citation ledger before you optimize anything — you can't manage the AI-answer game with a rank tracker, and the gap between "we rank" and "we're cited" is where all the insight lives. Second, give the agent real but bounded authority: let it ship the small reversible stuff relentlessly, and make every new page and every argument-changing rewrite wait for a human. The value is in the relentlessness of the small moves, not in autonomy for its own sake. Third, accept that this is a monitoring discipline, not a project. The platform moves, competitors refresh, and citations you won last month quietly leave. The team that reads the ledger every morning beats the team that runs a brilliant audit twice a year — not because they're smarter, but because they're present when it changes.

We didn't set out to write a manifesto about agentic SEO. We set out to stop doing the boring parts of our own marketing by hand, and the engine turned into the most honest research instrument we have, because it's pointed at us and it reports what's true whether we like it or not. The data above is what it's told us so far. If you want to see the system itself, the engine case study walks through it, and if you'd rather just talk about pointing something like it at your own site, that's a conversation we're happy to have.

Frequently asked questions

What is a citation ledger in AEO?

A citation ledger is a running record, per target query, of two things: whether an AI Overview appeared on that query, and whether your content was cited inside it. Tracked over a trailing window, it produces an AIO-present rate and a cited-when-present rate. It's the core metric for answer-engine optimization because it measures the thing that actually drives AI-assistant visibility — being in the answer box — which classic rankings and clicks no longer reliably capture.

Can an AI agent really run an SEO program autonomously?

Partly, and the boundary is the point. In our program an autonomous agent handles data pulls, AI-Overview citation audits, internal-link fixes, freshness and answer-block tweaks, re-index requests, and daily reporting — all the relentless, reversible work. It does not publish new pages or rewrite the argument of existing ones; those always wait for human review. The model that works is bounded authority: the agent ships small safe changes continuously and queues anything that changes meaning or can't be easily undone.

Why does my page rank but not get cited in the AI Overview?

Because ranking and citation are decoupled. Google assembles an AI Overview from whichever pages have the cleanest, most current, most liftable passage for each sub-claim — a selection only loosely correlated with classic ranking. A page can rank number one and be absent from the Overview above it, while a page outside the top ten gets cited. Winning citation is a separate workstream: tight answer blocks, FAQ structure, first-hand statements, and freshness, optimized for extraction rather than for blue-link position.

Did Google's 2026 "Expert Advice" change affect AI Overview citations?

In our data, yes and noticeably. Beginning in early May 2026, AI Overviews started surfacing forum posts and first-person experience much more aggressively. Our citation ledger recorded single-source practitioner citations being displaced by Reddit threads and discussion blocks on several queries within the same week — not because the content changed, but because the selection criteria did. It's the clearest example we have of why AEO requires continuous monitoring: the platform re-weights what it rewards without notice.

How do you measure AI-assistant traffic from ChatGPT and Perplexity?

Modern analytics now resolve a distinct "AI Assistant" channel that groups referrals from ChatGPT, Perplexity, Gemini, and Google's AI surfaces. It became legible in our analytics only recently, and it exists because of citations rather than classic rankings — which is why the citation ledger and this traffic channel move together. It's still small in absolute terms for an early-stage site, but it's the first directly attributable evidence that AEO work converts into sessions.