Double Jeopardy in AI Search

A marketing law says smaller brands lose twice: fewer buyers, slightly less loyal. In two months of ChatGPT and Google AI Mode answers to the same buyer questions, the second loss was far steeper, and the two multiplied into a sales channel that barely reaches the companies it rarely names.

On this page
  1. The Law Says a Smaller Brand Loses Twice, and Only Slightly the Second Time
  2. A Handful of Companies Take Most of the Mentions
  3. A Smaller Company Named Once Is Usually Not Named Again
  4. Each Buyer Asks Once, So the Two Penalties Multiply
  5. Specializing Is a Narrower Way Out Than It Sounds
  6. Hand the Daily Asking to an Agent and Read Two Numbers a Month
  7. How Qvery Works on the Questions Your Rivals Own
  8. Compare Your Reach With the Leader
  9. Frequently asked questions

If your company competes with one or two names everyone in your market already knows, marketing science described your position decades ago. You have fewer buyers than the leader, and your buyers come back slightly less often.

The common belief now is that AI assistants break it. A buyer describes the situation in full, the assistant matches on fit, and the best match gets named whatever its size. Owners of companies with a few hundred people hear this from agencies and vendors, and it decides where money and staff time go.

We checked that belief against repeated AI answers. At Qvery, we measure where buyers encounter your company and work on the pages those answers read; below, I explain how we take on that job.

The law's pattern holds, and more steeply than in buying. A company named on fewer of its market's buyer questions is also named less often each time one of them is asked again.

The two penalties multiply, limiting how often buyers see your company. These findings cover mainly business-software markets. Smaller here means named less, not fewer employees; a name in an answer is not a sale.

62.21%of all the answers to a market's buyer questions name its most-named company (median market, ChatGPT)
0.0435%name the typical company the answers mention at all, about one answer in 2,299
40.42%of the companies the answers name in a market appear in exactly one answer in two months (ChatGPT)
Source

Qvery's GEO Agent: answers to tracked buyer questions, ChatGPT and Google AI Mode, August to September 2026

The Law Says a Smaller Brand Loses Twice, and Only Slightly the Second Time

In their 1990 paper Double Jeopardy Revisited, Ehrenberg, Goodhardt, and Barwise describe the pattern: a small brand has fewer buyers, who also buy it less often.

The Ehrenberg-Bass Institute traces the law to the 1960s and finds it across consumer and business-to-business markets.

The institute describes firms with more market share as having many more buyers who are slightly more loyal.

That is why the law points to winning more buyers: loyalty follows size, in the institute's account.

AI answers carry the same two measures once you line them up. A buyer question stands in for a buyer, and the same question asked again on another day stands in for the next purchase.

A Handful of Companies Take Most of the Mentions

Start with the first penalty. In the typical market, the five most-named companies took 47.78% of every company mention on ChatGPT and 36.48% on Google AI Mode. The single most-named company took 13.25% and 12.81%.

The Leaders' Share of Every Company Mention in AI Answers

% of company mentions

ChatGPTGoogle AI Mode
View the data
LabelChatGPTGoogle AI Mode
The most-named company13.25%12.81%
The top five47.78%36.48%
The top ten63.32%51.66%
Median market: the share of all company mentions taken by the most-named company, the top five and the top ten, ChatGPT against Google AI Mode. Buyer questions Qvery's GEO Agent asked every day in August and September 2026, mostly business software; questions naming a tracked company are excluded.Qvery's GEO Agent: repeated answers to tracked buyer questions, ChatGPT and Google AI Mode, August to September 2026

These answers come from Qvery's GEO Agent, which asked the same buyer questions on ChatGPT and Google AI Mode day after day for the markets it tracked, mostly business software, in August and September 2026. Questions that named a tracked company are left out, and typical means the median.

The other end of the list is where the less-named company lives. Four in ten companies the answers named in a market appeared in exactly one answer in two months: 40.42% on ChatGPT, 43.75% on Google AI Mode. More than half appeared twice or less.

A Smaller Company Named Once Is Usually Not Named Again

For the second penalty, sort companies by penetration, then compare their repeat rates.

The Fewer Questions Name a Company, the Less Often Each One Names It Again

% of answers naming the company

ChatGPTGoogle AI Mode
View the data
LabelChatGPTGoogle AI ModeNote
Under 5%2.17%2.17%Median penetration in this band: 1.01%.
5% to 10%5.67%4.71%Median penetration in this band: 6.12%.
10% to 25%8.77%6.83%Median penetration in this band: 14.00%.
25% to 50%16.24%13.15%Median penetration in this band: 33.00%.
50% or more38.06%29.07%Median penetration in this band: 72.50%.
Median repeat rate (the share of all answers to a question that name the company, on the questions where it is named at all, averaged), by the share of its market's questions on which the company is named. ChatGPT and Google AI Mode, August and September 2026; companies from every market Qvery's GEO Agent tracked.Qvery's GEO Agent: repeated answers to tracked buyer questions, ChatGPT and Google AI Mode, August to September 2026

Inside every market we tracked, the companies named on a quarter or more of the questions had a higher repeat rate than the companies named on under 5%.

On a shop shelf the smaller brand's buyers are almost as loyal as the leader's. In an AI answer the smaller company rarely gets a second mention.

Part of that gap is arithmetic. A company named once in about 46 askings sits at 2.17% by construction, and companies named exactly once pull the bottom of the line down.

So we split the two numbers across months: penetration from August's answers, the repeat rate from September's answers to the same questions. The gap held. The typical company named on under 5% of the questions in August had a September repeat rate of 0.00% on both engines.

Most of them vanished outright: 69.30% were not named on those questions once in September on ChatGPT, and 73.97% on Google AI Mode. The typical company named on half or more had 39.52% and 31.11%, and none went unnamed.

The split does not remove the plainest explanation: an engine names some companies more readily than others, and that habit shows on both counts. In buying, that is the law's explanation too. The penalty comes with the smaller share, not with anyone's intent.

The two measures sit on different scales, so the comparison with buying is one of shape, not of a number.

Two dot grids, rows are buyer questions and columns are days asked. The leader is filled on four of five rows, about eight dots each; the smaller company on one row, two dots.
Read the rows, then the dots in each row: the smaller company loses on both.Illustration by Qvery, drawn for this article.

Read the drawing by rows, then by dots: fewer rows for the smaller company, and fewer dots on the rows it reaches. A buyer who asks any one of those questions, on any one day, is far more likely to read the leader's name.

Vlad ShvetsPartner @ Qvery

A screenshot of your company in one ChatGPT answer is one coin that landed heads. It cannot tell you whether buyers will see your name tomorrow. Ask for the same questions over several weeks before treating that screenshot as reach. The question for whoever sent it is how often it lands heads, and on how many questions.

Each Buyer Asks Once, So the Two Penalties Multiply

Gartner's survey of 645 business buyers supports that objection: buyers used seven sources on average during a recent purchase. 45% used generative AI, mostly for vendor and product information, and 69% preferred to check AI insights with a salesperson.

An answer is one step toward a shortlist. This data says nothing about who signed.

But a buyer who asks once gets one draw from that question's answers. Nobody knows which day's answer they will see.

So the repeat rate is that buyer's chance of seeing you on that question, asked the way we asked it.

Across a whole market, the chance is your penetration times your repeat rate. The product is your share of answers: the share of the answers to your market's questions that name you, each question counted once.

25.79%of all the answers name the typical company found on half or more of its market's questions
1.26%name the typical company found on 10% to 25% of them
0.0377%name the typical company found on under 5% of them
Qvery's GEO Agent: share of all answers to tracked buyer questions naming the company, ChatGPT, August to September 2026

From the top band to the bottom, the share of all answers falls 684.21 times on ChatGPT and 507.78 times on Google AI Mode, far more than either penalty alone.

Buyers do not ask every question equally often, so read this as the shape of the channel, not a forecast of your inquiries.

The objection has a second form: a company named on few questions will be named rarely on each of them by arithmetic alone. The law works the same way in buying, where loyalty, in the institute's words, "is largely a predictable function of market share."

That is what makes the law useful. If AI answers behave like buying, the repeat rate will not move on its own, and the bigger lever is the first number: more of the questions.

The institute's finding for brands that grew is the same, with winning new buyers roughly twice as important as keeping old ones.

Specializing Is a Narrower Way Out Than It Sounds

The usual advice is to specialize where fit should beat size. Two findings narrow that way out.

First, the specialist pattern is rare. The law has a known exception, the niche brand with few buyers who are unusually loyal.

Among companies named on under 10% of a market's questions, 0.63% on ChatGPT and 0.57% on Google AI Mode were named in half or more of the answers to the questions where they appeared.

Longer questions favored the leaders. We compared each market's shortest third of questions with its longest third, which included the buyer's situation and constraints. The pattern held after removing questions that named a leader.

Longer Questions Gave the Leaders More of the Answer

% of company mentions

Shortest third of questionsLongest third of questions
View the data
LabelShortest third of questionsLongest third of questions
ChatGPT: the most-named company11.41%16.15%
ChatGPT: the top five42.05%51.37%
Google AI Mode: the most-named company9.82%14.76%
Google AI Mode: the top five36.02%41.97%
Share of company mentions taken by the market's most-named company and by its top five, on the shortest third of each market's questions (typically five words) and the longest third (typically 32 words), median market. ChatGPT and Google AI Mode, August and September 2026.Qvery's GEO Agent: repeated answers to tracked buyer questions, ChatGPT and Google AI Mode, August to September 2026

We expected the opposite. These answers do not say why it happens, and they test no single specialist; they show that the longer questions in these markets did not open room for the smaller names.

The Ehrenberg-Bass Institute describes unusually loyal companies as often serving one region or sub-market, limiting their reach and putting a ceiling on growth.

If you serve one city or one kind of client on purpose, the ceiling comes with the choice. If you sell to a whole market and the answers name you only in a corner of it, they are showing you low penetration: the first penalty, the one the law points to.

Hand the Daily Asking to an Agent and Read Two Numbers a Month

If each answer were an independent draw, a company named in one answer in ten would go unnamed across five askings in a row 59.05% of the time. A person who checks a question now and then cannot tell a low repeat rate from none.

An agent can repeat the asking each day; your pipeline owner decides what the monthly numbers mean. Give the agent this brief:

  1. Build the Question List From Your Own Pipeline

    Take broad and specific buyer questions from sales emails, inquiry forms, and lost-deal notes. The pipeline owner approves the list, because it decides what gets measured.

  2. Ask Every Question Every Day on Both Engines

    Run the questions daily on ChatGPT and Google AI Mode. Keep the engines apart in every number because they name companies at different rates.

  3. Record Every Company in Every Answer

    Record your name and every other name per answer, with the pages it cites. Read your numbers against those competitors.

  4. Compute the Two Numbers Once a Month

    Calculate penetration, repeat rate, and their product per engine over a rolling two months. Compare them with the tables below.

  5. Hand a Person a Worklist, Not a Dashboard

    List questions where a rival appears in most answers and you in none, with the cited pages. The pipeline owner chooses which pages change and approves publication.

  6. Count What It Replaced

    Time one manual pass of the list on both engines with a stopwatch; the agent runs about thirty of those passes a month. Count errors too: each decision taken on one screenshot that the two numbers would have reversed.

The GEO Agent handles daily measurement; the custom worklist and monthly calculations belong to the brief above.

Read your repeat rate against the medians below. They describe the markets we tracked, not a promise for yours.

ChatGPT: typical repeat rate by penetration band, August and September 2026
Named onTypical, %Middle half, %
Under 5%2.172.13 to 4.35
5% to 10%5.673.44 to 10.11
10% to 25%8.775.01 to 15.49
25% to 50%16.248.72 to 26.42
50% or more38.0622.26 to 51.41

Google AI Mode runs lower than ChatGPT in every band above the first.

Google AI Mode: typical repeat rate by penetration band, August and September 2026
Named onTypical, %Middle half, %
Under 5%2.172.13 to 4.26
5% to 10%4.713.00 to 8.93
10% to 25%6.834.30 to 12.61
25% to 50%13.157.36 to 21.23
50% or more29.0715.72 to 44.42

Your penetration sets your band, and your repeat rate lands in one of three places:

  • In the middle half: you are typical for your band, and if AI answers follow the law the way buying does, the way up runs through more questions rather than a higher repeat rate on the ones you have.
  • Well above it: look for the corner of the market you own. In buying, the institute reads this pattern as a ceiling on growth, which may suit you if that corner is where your best clients ask.
  • Well below it: the answers name you on these questions but rarely, which is worth an afternoon of reading what they say about you before anyone decides why.

How Qvery Works on the Questions Your Rivals Own

Our GEO Agent shows where rival names keep appearing instead of yours. Your brand, competing companies, and industry inform the questions it creates. It runs them daily by country. ChatGPT and Google AI Mode supply the answers; the product retains their full text and citations.

Read the sentiment behind your mentions, check your visibility, and judge your share of voice against rivals by country and topic. We examine the cited sources with you, then work on the facts and pages the engines read.

For the recurring work, we learn how your pipeline owner reviews answers and assigns page changes. We configure a custom agent in your existing systems to prepare that worklist on its own, leaving publication with the owner.

We put the first version of this review into operation within two weeks. Your staff learn from us how to handle its worklist, and we tally hours saved and incorrect page assignments prevented each week.

Your question list gives us the boundary of that job.

Compare Your Reach With the Leader

Before the next AI-search report, do one multiplication. Take your penetration and your repeat rate on each engine and multiply them, then do the same for the competitor the answers name most.

The ratio between the two products measures the gap, and if AI answers follow the law, the first factor is where it closes.

AI answers do not level the field for the company they rarely name. They multiply its two losses, and the second loss was nowhere near the slight one the law finds in buying.

Frequently asked questions

Does the double jeopardy law apply to AI search?

The pattern held in the answers we measured: companies named on fewer buyer questions were also named less often when those questions were repeated. That relationship held in every tracked market, but it describes mentions in answers, not purchases or company head count.

What is a repeat rate in AI answers?

Ask the same buyer question every day and count how often the answer names your company, on the questions where it names you at all. That share is your repeat rate. A buyer who asks once gets one of those answers, so it is also the chance that buyer sees you.

How do I work out my company's share of AI answers?

Multiply penetration, the share of questions that name you at least once, by repeat rate on those questions. The product is your share of answers, close to what reports call AI visibility. The median market leader appeared in 62.21% of ChatGPT answers in our data.

Can a smaller company get recommended by ChatGPT over a bigger competitor?

On a given day, yes: answers vary, and smaller companies do appear. Across many askings the bigger names appear far more often, on more questions and in more of the answers to each. Count both numbers on a fixed list of questions before you read anything into one answer.

Do longer, more detailed buyer questions favor smaller companies?

Longer questions did not favor smaller companies in our data. On the longest third of each market's questions, the leader took 16.15% of ChatGPT mentions against 11.41% on the shortest third; Google AI Mode moved the same way. The answers do not establish why.

How often should we ask the same question before trusting the result?

Ask daily for weeks, keeping the questions fixed. If answers were independent draws, a company named in one answer in ten would go unnamed across five askings 59.05% of the time. A handful of checks cannot reliably distinguish a low repeat rate from none.

Is being named in an AI answer the same as being cited?

A citation is a page listed as a source; a mention means your company appears in the answer's text. This article counts company names because those are what buyers see, and a cited page is a different thing to measure.

Who in the company should own this?

The pipeline owner should own the question list, because it comes from sales emails, inquiry forms, and lost-deal notes and affects what the company publishes. An agent can handle daily asking and counting; that person reads the monthly result and approves changes.

What does penetration mean in AI search?

Penetration is the share of your market's buyer questions on which an answer names your company at least once during the measurement period. It measures how broadly you appear across questions, while repeat rate measures how consistently you appear when those questions are asked again.

Does a high repeat rate mean our niche can grow?

A high repeat rate on a narrow set of questions may reflect a corner of the market you already own. The Ehrenberg-Bass Institute describes unusually loyal companies as often having restricted reach, so check whether that corner fits your intended market before treating it as a growth opportunity.

Can AI mentions predict how many inquiries we will receive?

The share of answers in this article weights each question equally, but buyers do not ask every question equally often. It describes the channel's shape rather than forecasting inquiries, and being named does not show who eventually buys.

What should we do when our repeat rate is below our band?

Read the answers that name your company and the pages they cite before deciding what is wrong. A low repeat rate tells you that mentions are inconsistent within that band; it does not establish the cause or tell you which page change will fix it.

Keep Reading

Property Managers Should Start AI With Maintenance, Not Leasing

Residents rank repairs just behind rent among their reasons to leave. Put a property company's first AI agent on the follow-through around each repair.

Brokers and Carriers Hang Up on AI Voice Bots

Carriers say they hang up on brokers' AI voices. The work worth automating is the call that only reports a fact, and most of those need no call.

AI-Drafted Work Doesn't Belong on the Billable Hour

Once AI drafts, an hourly bill carries only the review, and clients' own AI keeps adding rounds of it. A fixed fee with the rounds written in prices both.