Focus: AI search citations (AEO)
🔴 Impact: High
🟠 Difficulty: Mid
Hey {{First Name | there}},
Let's talk about the signals that decide whether AI search cites you.
Google denied three of them for a decade. Two turned out to be real.
I went back to the primary sources on this one, because the commentary going around has gotten sloppy. And along the way I'll hand you the free tool I built to score your own site against all of it in about 5 minutes.
I built that tool for a selfish reason: my Claude bill. I audit low-volume client sites, and the first version ran every page through an LLM. It worked, it cost more than those sites could justify, and it handed me different answers on the same page from one run to the next. So I rebuilt it deterministic. Anything a parser can settle, a parser settles. Same site, same numbers, every run. And the twelfth client site costs exactly what the first one did: nothing.
Google denied 3 things. 2 were real.
For most of the last decade the public guidance was consistent:
Clicks aren't a ranking factor
There's no site-wide authority score
There's no sandbox for new domains
Then the record caught up.
2023 antitrust trial. Google's VP of Search testified under oath about Navboost, which re-ranks results using aggregated click data on a rolling 13-month window. He called it one of the important signals. That is a long way from "clicks aren't a ranking factor."
The Content Warehouse leak. The schema holds a score for how tightly a site sticks to one topic, a measure of how far each page drifts from that centre, and a standard deviation of quality across the whole site. Google evaluates sites, not just pages.
The sandbox. This one actually held up, which I did not expect. More below.
💫 Here's the part almost nobody reframes. AI Overviews and AI Mode don't write answers out of nothing. They retrieve from Google's index first, then generate from whatever came back. Retrieval is the gate.
So these aren't ranking signals. They're citation signals. And three linked sources is a much smaller prize than ten blue links.
The 5 signals that decide if you get cited
Topical focus. siteFocusScore and siteRadius measure how tightly your pages cluster around one subject. This is the only signal on the list where a five-page site competes on equal terms. You can't out-authority a ten-year-old competitor. You can out-focus one this quarter.
Quality spread. siteQualityStddev is a standard deviation, not an average. Google stores how inconsistent you are. Which means deleting your weakest quartile moves the number more cheaply than writing ever will.
Estimated effort. One field, nine words of documentation: "LLM-based effort estimation for article pages." Everything else you have read about what it measures is guesswork. But a model estimating effort will find nothing to reward in a page that restates what five others already said.
Information gain. This is the one that annoys me, because it gets quoted constantly by people who clearly haven't read it. Everyone cites US11354342B2 as "Google rewards unique insight." Read the actual claims and it scores novelty against what that reader has already been shown, not against the internet. Which is exactly how a three-source AI answer gets built. The third source is valued only for what it adds to the first two. Redundancy scores zero.
Sustained referencing. Not your backlink total. Whether people still point to you now. A link profile that flatlined years ago tells Google you're a source that mattered once and then stopped.
⚠️ The denial that held: hostAge gets quoted at half length. Read it whole and it says the data is used to "sandbox fresh spam." Fresh spam, not fresh sites. On this evidence Google was telling the truth, and new sites can be cited.
Do these 4 things this month
Every one of these maps to something Google actually stores. The catch is that you can't eyeball any of them on a 200-page site, so I'll point at where each one turns up in the audit as we go.
Prune before you publish anything new. siteQualityStddev is a standard deviation, so your worst pages drag the whole site down and deleting your weakest quartile moves that number more cheaply than writing does. Knowing which pages are the weakest quartile is the hard part. That's the per-page report and the duplicate clusters tab: every URL scored, near-identical pages grouped together, and a keep recommendation for each cluster.
Pick one subject and stay inside it for a full quarter. This is siteFocusScore and siteRadius. The pages hurting you are the ones sitting furthest from your site's centre, and most people can't name them off the top of their head. The topical outliers tab ranks them by distance, so your pruning list writes itself.
Put one non-reproducible thing in every page. A screenshot, a number, a test you ran. This is information gain, and it's the cheapest, highest-lift fix in the set. It's also the one nearly everyone fails. The first site I ran this on had static schema on 100% of pages and an authoritative citation on 1.3%. The fix queue flags every page with no cited source and no attributed statistic, ranked by lift per hour, so you start with the pages where it pays most.
Stop changing dates in titles. The 2003 patent explicitly discounts cosmetic edits, so bumping the year without touching the body buys you nothing. The audit compares dateModified against datePublished and flags the pages where the date moved but the content didn't.
All four of those live in the same place. Paste one URL into a free Google Colab notebook, hit Run all, and about 5 minutes later it has crawled your whole site, scored every page against the ten signals that decide retrieval, and dropped the formatted Google Sheet in your Drive. A Google account is the only other thing you need. No API keys, nothing to install, and no LLM calls in the scoring, which is why running it again next month, or on the next client site, costs you nothing.
💫 Every check is tagged with where its evidence comes from. The leak tells you what Google stores. The Princeton GEO study tells you what actually moves citation rate. Those are two very different kinds of evidence and the sheet doesn't pretend otherwise.
⚠️ One honest caveat. contentEffort has nine words of documentation and no stated inputs. Any tool claiming to score it is inventing a number, mine included. So I left it out.