The GEO Metric Nobody Can Show You (And What We Found When We Went Looking)
AI answer engines rarely say where their citations come from. We stopped guessing and checked ours by hand. Here is the method, and what it cannot tell you.
- → Why the Search Console for generative engine optimization arrived in June and still can't tell you much
- → What happened when we ran it by hand: eight questions, one citation
- → The part that surprised us: who won the other seven
- → A method you can run this week, and the four things it can't tell you
In Comedy You Know Immediately
I was doing well at New York Comedy Club 24th St. This woman's laughing like crazy, and I want to keep her going so I go "you sound like a dying pelican." Everybody laughs hard, except her.
She shouts: "I've been laughing at all your jokes and you make fun of me? I'm done laughing."
She was. And so, it turned out, was the room. I tried explaining it was a joke, which is the comedy equivalent of performing CPR on someone at their funeral.
That's stand-up. You never wonder whether it's working. The answer arrives in under a second, out loud, whether you asked for it or not.
I've run four businesses before Ingenium Vector. Health insurance telesales in Chicago. My family's health food store in Florida. A natural products sales firm in the Northeast. My stand-up career. I always knew immediately. The store reorders or it doesn't. The guy on the phone signs or hangs up. You count the drawer at close and the day has a number. Who the couple on a first date was.
Not in this business. A client just asked if their GEO was improving and I did the Homer Simpson into the bush meme.
First, What Our Own Instruments Actually Do
Before telling anyone how to measure this, I checked what we can measure. As of August 2026:
| What we want to measure | What we actually have |
|---|---|
| AI answer-engine citations | Partly, since June 2026. Search Console now shows impressions inside Google's AI Overviews and AI Mode: that you appeared, not what it was worth. No clicks, no queries, no position, a subset of sites, Google only. Perplexity, the engine we tested on, publishes nothing |
| Search impressions and positions | One hardcoded day per run, so we hold about three separate days a month, not a series. Chart them as a line and it reads as a cliff: the days between aren't zero, they're absent |
| Crawl and index status | Nothing |
| Schema validity on live pages | Checked at ship. Not monitored, and I'm not going to write the word monitored on a thing we look at once |
One of the four arrived in June, half-built. The other three are missing, crippled, or checked once and called monitoring. That's why the honest answer to "is it working" was a shrug rather than a lie.
I've Paid For Beautiful Numbers Before
Years ago I ran Google ads for my sales firm's site, five dollars a month against "Omega 3." The cost per click was unreasonably good. I remember being pleased with myself.
The problem was that I didn't sell anything on that website.
I was a middleman for boutique manufacturers on one side, health food stores on the other. Nobody could buy a bottle of fish oil from me. I was paying strangers to land on a page and be confused.
That worries me more than having no number at all. A number that's easy to get will quietly become the number you manage.
So We Did It By Hand
The metric the 2026 guidance keeps landing on is called Share of Model. Before using it I should say that the field can't agree what it means, which is its own kind of answer.
Some define it the way I'm about to: of the AI answers to your buyers' questions, what fraction mention you at all. Others make it relative, your mentions over all brand mentions in the category, and would call mine a mention rate. Both are in active use as of 21 August 2026, and at least one source uses Share of Voice and Share of Model interchangeably.
I use the first. So what I measured is a mention rate with no competitor denominator, and I'd rather say that than let the term do quiet work.
Outside that Google report, the answer engines publish nothing official. Third-party trackers sell this, and what they sell is the manual method below run at scale. True as of 21 August 2026, and the most perishable sentence here.
So on 18 August 2026 we ran it by hand for our client The Law Office of Jason Browne PLLC, a Brooklyn tenant-side housing practice. Eight questions a real tenant would type. One run each, on Perplexity, full answer text read for any citation back to the firm.
Result: cited on one of eight.
The one hit was the hire-a-lawyer question, some version of "Brooklyn eviction defense lawyer." Perplexity returned a four-option table of private counsel and the firm was one of the four. One run, one engine, one day, no location controls, so that's a fact about what a search returned and not a claim about where the firm stands among lawyers. That table sat underneath a section about free legal help.
The Part That Surprised Us
I assumed the other seven went to competitors. They didn't.
They went to nyc.gov, the state housing agency, the courts, Legal Services NYC, Legal Aid, Housing Court Answers, Met Council. Government and nonprofit almost all the way down. On informational questions the engine treats the government as the source and law firms as a directory it appends at the end, if at all.
Two things stopped that being a tidy story.
One private firm did earn an informational citation, about HP actions. Somebody got through.
And the two clearest openings were questions where our client already owns a matching tool, a rent-stabilization quiz and an overcharge calculator, and still didn't surface. The asset exists. The engine hasn't connected it to the question. That's fixable, and we wouldn't have known.
We aren't the only ones seeing this. Pew Research Center tracked 68,879 real Google searches in March 2025 and found .gov sites cited inside AI summaries at three times their share in standard results. Directional, not backup: that's Google's AI Overviews, not Perplexity, the data is seventeen months old, and Google pushed back on the framing. The same study found people clicked a source inside the summary on 1 percent of visits. A citation isn't traffic.
If that holds in your category, and test it before believing it, the job isn't to outrank a competitor. It's to become what the engine reaches for after it finishes quoting the official source. Different job, different content, and most GEO advice hasn't run the check.
What This Does Not Prove
This is the part most people skip, so here it is in the middle where you can't.
- One unrepeated run. One of eight is a single observation, not a rate. Run it Thursday and it might be two, or zero.
- No personalisation or location controls recorded, and both plausibly matter.
- No prior series, so no trend. I can't tell you it improved. There's no earlier number to improve on.
- One engine. Perplexity and nothing else, which makes this a Perplexity finding rather than an AI finding. Do as I say, not as I did last Tuesday.
What it's good for is direction: what kind of gap, and where. That beats a percentage with a decimal point.
Run It On Yourself This Week
Freeze ten questions a buyer would type when they're close to hiring someone like you. Not category questions. Hiring questions, and the informational ones just before.
Run each across several engines, because they pull from different places and rarely agree.
For each, record three things: mentioned or not, described accurately or not, and who got cited instead. That third column is the one that pays.
Then date it and repeat monthly. The first run is a baseline, not a verdict. Ninety minutes the first time, half that after, which is the same arithmetic we run on any automation before we build it.
The Thing That Actually Kills This
Not the result. The friction.
For years I wouldn't look at my credit card bills. Not fear of the balance, just a sequence of small dull steps my brain wouldn't start. Then I put them on autopay and it stopped being a problem. Nothing about the money changed. I removed the monthly decision.
Your GEO check dies the same way, and not from a bad result. It dies because month three arrives and running it is forty-five minutes of small dull steps nobody put on a calendar.
So put it on one. Same day each month, an actual name against it, with the ten questions already written down. And decide who owns it when it comes back bad, because if the answer is nobody, that's the same failure as an automation with no owner after it breaks, and it ends the same way.
If you froze ten questions for your own business right now, which one are you most afraid to run?
Smatthew Cohen is an AI Operator and the founder of Ingenium Vector. Before that he ran a sales firm called Tortoise & Rooster for twelve years, helping boutique manufacturers who couldn't afford the agencies that were ignoring them anyway. He builds things now.