Your positioning statement is no longer the one on your homepage. It is the one an assistant recites when no one is watching.

A one-hour diagnostic that costs nothing and will probably ruin your afternoon.

1. The paragraph you did not write is doing your selling

Somewhere this morning a buyer typed your category into a chatbot and read a paragraph about your company that you have never seen. She did not visit your website. By the time she books a call the shortlist is drafted, and you are on it or you are a name she scrolled past. The numbers are not soft:

1.  Forrester’s 2026 Buyers’ Journey Survey of nearly 18,000 business buyers found 94% used generative AI in the purchase process, up from 89% a year earlier, and twice as many named gen AI their most meaningful research source as named any other.

2.  G2 surveyed 1,076 B2B software buyers in March 2026: 51% now begin research in a chatbot rather than a search engine, up from 29% a year earlier, and 71% use AI search tools specifically for vendor research.

3.  69% of those buyers chose a different vendor than planned because of what a chatbot told them, and 33% bought from a company they had never heard of.

4.  Semrush polled 519 US B2B professionals who use AI at work: 92% said AI has shaped their vendor shortlist, 45% significantly, and only 7% notice a vendor in an AI answer because they recognize the name.

5.  6sense found the eventual winner was already on the buyer’s Day One shortlist 95% of the time. Apollo puts the average 2026 shortlist at roughly 2.5 vendors, down from 3.2.

Read the last two together and the arithmetic gets uncomfortable. Recognition buys almost nothing at the moment of the ask, the list has about two and a half chairs, and the seating chart is drawn before you know a meeting is happening.

2. The protocol: five machines, three questions, one hour

Open five assistants in five fresh sessions and log out of all of them: no memory, no custom instructions, no prior turns. A logged-in session flatters you, and a model sounds better once you have already discussed the company in that thread, which is not what a buyer sees. Paste all fifteen responses into one document verbatim.

Table 1. The Mirror Test, start to finish

Step	What you do	Why it matters
Question 1	“Describe [company] in two sentences.”	This is your positioning statement, as recited by the machine. Compare it to the one in your deck.
Question 2	“Who are [company]’s main competitors?”	This is your category. If the machine puts you in the wrong one, price and pitch stop making sense.
Question 3	“Who should hire [company], and who should not?”	Your qualification criteria. The second half is worth reading twice.

3. The variance is the finding

A company with a real position gets five near-identical paragraphs. A company without one gets five different companies, and at least two of them belong to somebody else. One load-bearing caveat before you panic at a single bad answer: a July 2026 arXiv paper fit a variance model to 12,933 responses about 20 brands, and brand identity explained 1.5% of the variance in one answer. Reliability from a single screenshot came out near 0.01, a statistician’s way of saying you have learned nothing.

That same paper tells you how to spend hour two. Paraphrases and extra models buy reliability; repeats of an identical prompt do not. Fifteen paraphrases run once scored 0.347, while five paraphrases run five times each scored 0.332 and cost 40% more queries. Separately, only 2.2% of cited sources stayed identical across three repeated runs, with week-to-week citation shifts of 56% to 74%.

Table 2. How to grade what comes back

Verdict	What you are looking at	What it means on Monday
Convergent	Five paragraphs that differ in phrasing and agree on substance. Same category, overlapping competitor sets, same buyer.	You have a position. Protect it, check it quarterly, spend your afternoon elsewhere.
Drifting	Same category, but competitor lists barely intersect and the “who should hire” answers point at different budgets.	Your category is legible and your differentiation is not. That is a comparison-page problem.
Fractured	Three or more materially different descriptions. One flatters, one is generic, one paraphrases your 2021 tagline.	Your public record is stale and thin. The machines are averaging old sources because nothing recent outranks them.
Mistaken identity	A model confidently describes a company with a similar name, or attributes a competitor’s product to you.	Stop everything else. A wrong fact repeats across every answer that inherits it.

4. Why the five disagree in the first place

They disagree because they are reading different internets. Writesonic examined 161,286 prompts across four engines in May and June 2026: the four shared roughly 17% of cited sources on the same prompt, and only 3.8% were cited by all four.

Table 3. The five machines have five reading habits

Engine	Its favorite source type	Citations per answer	Overlap with Google’s top 10
ChatGPT	Wikipedia (47.9% of top citations)	roughly 6	6% to 8%
Perplexity	Reddit (46.7%)	21.87, the highest of any engine	28.6%, the outlier
Google AI Overviews	YouTube (23.3%)	8.34	about 38%, down from roughly 76% in mid-2025
Claude	Independent blogs (43.8%)	5.67	low; Brave-based retrieval

Ahrefs studied 15,000 prompts and found only 12% of cited URLs rank in Google’s top 10 for that query. An analysis of 250 million AI results found classic SEO metrics explained 4% to 7% of citation variance. Your rankings are not a proxy for this.

5. What Pamplona knows about packs, and you do not

Every July at eight in the morning, six fighting bulls and six steers run 875 meters through the old quarter of Pamplona in about two and a half minutes. The city sets 2,000 planks and 300 posts of fencing starting in early June.

The steers, the mansos, are not decoration. Their job is to keep the herd moving as one animal. The most dangerous outcome on the course is a suelto, a bull separated from the pack, which becomes disoriented, unpredictable, and may turn and run back the way it came. A herd running together is fast and legible. One animal running alone puts people in the hospital.

Table 4. The encierro, and the diagnostic

On the course	The record	In your five answers
The herd runs together	A clean run covers 875 meters in 2:14 to 2:28. Fastest on record: 2:05, July 14, 2015.	Five convergent paragraphs. Boring, fast, survivable. This is what you want.
A suelto breaks away	July 11, 1997: a 600-kilo Jandilla bull named Huraño ran the route alone in 1 minute 45 seconds, far ahead of the rest.	The one model describing a different company. Faster and more confident than the others, and wrong.
The pack refuses to move	In 1959 a Miura run took a full half hour because one bull would not enter the pen. A sheepdog solved it.	Every model hedges and refuses to commit. You have given them nothing specific enough to say.

The Associated Press keeps noticing the same detail year after year: many runners appear completely unaware when bulls are breathing down their necks. Sixteen people have died at San Fermín since record-keeping began, most recently in 2009, and the morning injuries are overwhelmingly falls suffered by people who never looked behind them. Running the Mirror Test is looking behind you.

6. Trace every wrong sentence back to the page that fed it

Now do the archaeology. For each wrong claim, ask the model where it got that, then read the page. It is usually not your page. Muck Rack found 84% of AI citations come from earned media rather than a brand’s own site, and OtterlyAI found 66% of B2B software recommendations happen without ChatGPT citing the brand’s site at all. The field calls that a ghost citation.

Table 5. Symptom, source, remedy

What the machine said	Where it almost certainly came from	What you actually fix
An old tagline or a dead product	A release, conference bio, or directory listing that outranks your homepage in the retrieval layer	Update the third-party record first. Your homepage is not the page that fed the answer.
The wrong competitor set	A listicle or review-site category page that filed you under the wrong heading	Get recategorized on the review sites. G2 names review-site citations the top trust signal for buyers reading an AI recommendation.
A competitor’s feature credited to you	A comparison page, yours or theirs, written loosely enough to blur	Rewrite the comparison page with explicit attribution in every row.

The payoff is measurable. Stacker found earned-media distribution lifted AI citations by a median of 239%, Ramp reported a sevenfold rise in AI mentions within one month, and cited brands see roughly a 23% lift in branded search over 30 days even when click-through is negligible.

7. A focus group with a 100% response rate and no invoice

Consider what you would normally pay to learn what the market thinks you are. One in-person focus group runs $7,000 to $12,000 after recruiting, facility, moderator, and incentives. Three or four groups run $25,000 to $75,000. An annual tracker across three or four markets runs $50,000 to $120,000.

The Mirror Test has no moderator to lead the witness, no screener to bias the sample, and no incentive buying agreeable answers. One limitation is worth stating out loud: it measures what the machine believes, and machines are not customers. That shrinks every quarter, because 94% of your buyers walk through the machine on their way to you.

8. Rewrite the page the machine read

Here is the twist most teams get backward. Faced with five wrong descriptions, the reflex is to polish the homepage, because the homepage is the page you are proud of. It is frequently not the page that produced the answer. Fix the source, in this order:

1.  Correct the wrong fact wherever it originated, even if that is a directory entry, a stale release, or a speaker bio from 2023.

2.  Fix your category placement on the review sites before you touch a word of website copy.

3.  Write the comparison page you have been avoiding, the one naming competitors that says plainly who should not hire you. Averages have nothing to quote.

4.  Publish numbers. Retrieval systems can only cite claims that exist in citable form. Adjectives are not citable.

5.  Re-run in 30 days, varying the wording rather than repeating it, and log the drift. You are watching a herd, and herds move.

Only 14% of organizations track AI citation visibility, while 43% call AI optimization a core 2026 strategy. That gap is your window, and the fencing goes up in June whether you are ready or not.

Sources: Forrester Buyers’ Journey Survey 2026 (n≈18,000); G2, The Answer Economy, March 2026 (n=1,076); Semrush B2B AI survey, March–April 2026 (n=519); 6sense 2025 Buyer Experience Report; Apollo 2026; Writesonic citation-overlap study (161,286 prompts, May–June 2026); BrightEdge; ZipTie; Ahrefs (15,000 prompts; 78.6M queries); Profound (30M+ citations); Qwairy (118,000 AI answers); Zatuchin, arXiv:2607.13304, July 2026 (n=12,933); Graphite, May 2026; Indig via Search Engine Land, June 2026; Muck Rack, May 2026; OtterlyAI, April 2026; Stacker, March 2026; GoodFirms 2026 SEO survey; Preuve AI and Drive Research market-research pricing, 2026; AP and CBS News San Fermín coverage, July 2026; sanfermines.net encierro records; University of Navarra Hospital.