Back to Blog

Absent, Wrong, or Second: Diagnose Your AI-Search Problem First

Seventy-two collected answers, twelve pages of approved budget, and the sorting exercise that moved the money to where the problem was.

Three matching navy letter trays with brass edges on a deep navy desk, photographed from a high angle. The left tray overflows with blank cream paper slips while a hand in a navy shirt cuff drops one more on the heap; the middle tray holds a few slips and an amber pencil; the right tray holds a couple of slips. Headline: Absent, Wrong, or Second? Sub-line: Sort the answers. Fund one fix.
Direct answer

Why does fixing the wrong AI-search problem waste budget?

Dr. Jonah Tebaa argues that "AI-search visibility" covers three different problems: the assistant does not name you (absent), names you with an incorrect or outdated fact (wrong), or names you correctly but puts a rival first (second). Each needs a different remedy and a different owner, so money spent on the wrong one changes nothing. His Three-Verdict Sort gives every collected answer one verdict, funds only the largest group for 30 days, and re-runs the same questions afterwards.

A 40-person firm had approved the budget for twelve rewritten web pages. When they finally sorted the AI answers they had collected, 9 of the 72 were about the problem those pages would fix. The other 63 were something else.

The firm in this piece is a composite, built from patterns in my work. It is not a real client, and every number is illustrative, not research. I use it because the decision it describes is one I see owners face regularly: a vendor says "you need to be visible to AI", a budget is approved, and nobody has yet established what is actually wrong.

One Complaint, Three Different Problems

When an owner says "the AI gets us wrong", the sentence can mean three different things. Sometimes the assistant never mentions the business. Sometimes it mentions the business and states the pricing from two years ago. Sometimes it describes the business accurately and then recommends a competitor first. In each case the owner feels the same discomfort, and in each case the cause sits somewhere else.

I have come to treat these as three separate problems with three separate remedies. Being left out is a question of whether you exist in the places assistants draw on. A wrong fact is a record-keeping question. Being second is a question of proof: the rival has published evidence that you have not. A rewritten website can touch the second and third only indirectly, and the first not at all if the pages are the only thing you control.

That is why a single vendor remedy, applied to all three, usually disappoints. My claim is simple: most businesses pay to fix an AI-search problem they have not diagnosed. The cheapest step available to them is to sort what the assistants already say before approving anything.

The Three-Verdict Sort

The method has a fixed order. First, fix a panel of 20 to 30 questions a real buyer would ask: category questions, problem questions, comparisons, and "best for my situation" questions. I described how to build that panel in Before You Trust an AI-Visibility Score, Build the Question Set, so I will not repeat it here.

Second, run each question in two or three of the assistants your buyers use. Save every answer as it came, with the date. Repeat a small sample a second time, so you learn how much the answers move by themselves.

Third, give each answer exactly one verdict. This is the practical core of the method:

  • Absent. You are not named at all. The remedy is earned presence: be mentioned in the places the assistants draw on, and publish one plain page that states what you do and for whom. Owner: the marketing lead or the owner.
  • Wrong. You are named, but a fact, an offer or a category is incorrect or out of date. The remedy is reconciling your facts across every place they appear, which I covered in Twelve Facts, Six Places. Owner: operations.
  • Second. You are named correctly, but a rival is named first, recommended, or described more strongly. The remedy is proof: named results, specific scope, a clear statement of who you are for. For comparison questions in particular, see The Comparison AI Assistants Quote Probably Isn't Yours. Owner: whoever holds the client proof.

Also count the answers where you are named first and correctly. Without that group the percentages overstate the trouble, and the denominator stops being honest.

Fourth, find the largest group. That verdict gets this month's budget, and only that one. If there is a tie or no clear leader, I choose the verdict closest to revenue, which is usually Second on the "best for my situation" questions. Fifth, run the same panel again after 30 days. The decision rule I state plainly to every owner is this: never fund two remedies in the same 30 days.

A Composite: 72 Answers, 12 Pages, and What the Sort Changed

Take the composite firm: 40 people, a Gulf-serving B2B software and implementation business. A vendor has recommended rewriting the website for AI, scoped at 12 pages. Before approving it, leadership collects 24 buyer questions and runs each in three assistants, which gives 72 answers.

The baseline sort came out like this: Absent 31 (43%), Second 22 (31%), Named first 10 (14%), Wrong 9 (12%). The total is 72. The same 24 questions run twice in one assistant differed on 6 answers, so I would treat any change smaller than about 6 as noise.

The tension is in the room before anyone opens the spreadsheet. The loudest complaint was "the AI has our pricing wrong". That was real, and it was also the smallest group: 9 of 72. The 12-page rewrite would have been aimed at the smallest bucket, while the largest one, being absent, would not have been reached by pages on the firm's own site.

So the firm funded Absent only, for 30 days. Three earned mentions in places its buyers' sources already cite: an industry association member page, a contributed piece in the trade press, and a partner's case write-up. Plus one plain page saying who the firm serves and what it does. The cost was about 30 working hours, not twelve rewritten pages. The pricing complaint was logged and passed to operations as a separate item for a later cycle.

The re-run after 30 days: Absent 22, Second 24, Named first 18, Wrong 8. The total is again 72. Absent fell by 9, which is larger than the 6-answer noise band. Second rose by 2, which is inside the band, so I would flag it and not celebrate or panic. Named first rose by 8, which also clears the band.

With Absent reduced, Second is now the largest group at 24, though only 2 ahead of Absent, which is inside the 6-answer noise band, so the tie rule applies and the verdict closest to revenue, Second, gets month two. The month-two remedy is proof: two named client outcomes with their scope, and nothing else that month.

Reading the Re-Run Without Fake Precision

The noise band is the part owners most want to skip, and it is what keeps the exercise honest. Assistants do not give identical answers to identical questions. If you re-run a sample and 6 answers change on their own, then a movement of 4 after your remedy tells you very little, while a movement of 9 tells you something. It still does not tell you everything.

The limits of the composite are worth stating directly. A panel of 24 questions is small. The assistants change between runs and between months. One run is a photograph, not a trend, and two runs are a comparison, not a proof of cause. Other things happened in those 30 days that nobody logged. The sort tells you where to spend. It does not prove that the spending caused the change, and I would distrust anyone who presents it that way.

The one-remedy rule is what makes even that modest reading possible. If the firm had also rewritten the pricing pages and published two case studies in the same month, a drop in Absent could not be attributed to anything.

What to Ask a Vendor Before Approving the Budget

Before you sign, ask for four things. Which verdict does your proposal address, and what share of my answers does that verdict represent? What is the noise band on the answers you are measuring? What will you change in the first 30 days, and what will you leave alone? And who on my side owns the work, given that the three remedies sit with three different people?

A vendor who cannot name the verdict is selling a general remedy for a specific problem. That does not make the work worthless. It means you cannot yet tell whether it is aimed at your largest group, and the sort costs a few days of someone's time to find out.

If you try this, I would like to know which of the three was your biggest group, and whether it matched the complaint you started with.

Related evidence: The paper introduces Generative Engine Optimization as a framework intended to help content creators improve how visible their content is in the responses generated by generative search engines. (the arXiv paper introducing Generative Engine Optimization)

Google's crawler documentation states that the Google-Extended control does not affect a site's inclusion in Google Search and is not used as a ranking signal there. (Google's documentation on the Google-Extended crawler token)

Frequently asked questions

What is the difference between being absent, wrong and second in AI answers?

Absent means the assistant does not name your business at all. Wrong means it names you but states a fact, an offer or a category incorrectly, or out of date. Second means it names you correctly but puts a rival first, recommends the rival, or describes the rival more strongly. Each answer gets exactly one verdict, because each verdict points to a different fix and a different owner.

How many questions do I need before the sort means anything?

I start with 20 to 30 questions a real buyer would ask, run in two or three assistants. That gives somewhere between 40 and 90 answers, enough to see which verdict is the largest group. It is a small sample and I treat it that way. The sort shows where to spend first. It does not measure your market share, and a larger panel is worth building once the first cycle has paid for itself.

Why fix only one verdict per 30 days?

If you fund two remedies in the same month and the numbers move, you cannot say which one moved them. One remedy per cycle lets you attribute the change, and it keeps the work inside the capacity of a small team. Answers also move on their own between runs, so a single remedy that clears the noise band is far easier to trust than two that overlap.

Can I run this sort myself without buying a tool?

Yes. A spreadsheet with one row per answer is enough: question, assistant, date, the answer as it came, and one verdict. Ask the questions in each assistant your buyers use, save the answers unedited, and repeat a small sample to see how much they move by themselves. A tool can save time later, but it should not replace the habit of reading the answers yourself.

Written by Brian, Dr. Jonah Tebaa's AI partner, on his behalf.