Ask an assistant what your company does. Then ask two more. If the three answers describe different companies, the fix is not a new homepage. Your homepage is one vote among hundreds, and right now it is losing the vote.
This is the part of AI search that catches marketing teams off guard, because it inverts twenty years of how positioning worked. You used to write the description, publish it, and everything downstream copied it eventually: analysts, journalists, your own sales deck. A model does not work downstream from you. It reads everything at once, weighs each source by how reliable that source has been on similar questions, and treats your own site as an interested party with a motive to oversell.
What the model is actually reconciling
For a typical B2B company, the material an assistant draws on looks something like the table below, and only the first row is under your direct control.
| Source type | What it usually says | How much weight it carries |
|---|---|---|
| Your website | Current positioning, aspirational category | Low to moderate, treated as self description |
| Review platforms | Category assigned by the platform, often years old | High, structured and cross checked |
| Listicles and roundups | Whatever the writer inferred in twenty minutes | High, frequently quoted |
| Communities and forums | How practitioners actually describe you | High for credibility, unpredictable in tone |
| Press and funding coverage | The description from your seed round | Moderate, durable, rarely updated |
You can rewrite your homepage in a week. You cannot rewrite what the internet decided you are without going and changing it, page by page.
The failure modes, in order of how common they are
- Stale category. You repositioned eighteen months ago. Every third party source still describes the old product, so the model does too, and you never appear in shortlists for the market you now sell to.
- Split identity. Half the sources call you a platform, half call you a service. The model resolves this by describing you in language so general that no buyer recognises a fit.
- Absent from the comparison set. Competitors appear in the roundups that get quoted for your category and you do not. Nothing on your own site can fix this.
- Correct but unquotable. Everything written about you is accurate and entirely generic. There is no specific claim worth lifting, so the model summarises the category and names someone else.
Measuring this with a worked example
Worked example
A repositioning that never left the homepage
A mid market software vendor moved from an SMB tool to an enterprise platform eighteen months ago. New pricing, new logo, new homepage copy. The team ran a baseline of three prompts across three assistants asking what the company does, who it is for, and who it competes with.
- Answers naming the old SMB category
- 7 of 9 responses
- Answers naming the current enterprise category
- 2 of 9 responses
- Sources cited when the model got it wrong
- Two review profiles last updated 3 years ago
Result: The team updated the two stale review profiles and emailed one roundup author with a correction. Re-running the same nine prompts six weeks later produced 6 of 9 correct answers, without a single change to the website itself. The homepage was never the problem. The two profiles nobody had logged into since the old pricing page went live were.
The repair sequence
- 1
Establish the current description
Run the same three prompts across the major assistants: what does this company do, who is it for, who competes with it. Record the answers and every source they name. This is your baseline and it takes an hour.
- 2
Write one canonical description
Two sentences. Category, buyer, the specific thing you do that others do not. This exact text goes on your site, every profile, every partner page, every conference bio and the boilerplate of every press release. Consistency is doing most of the work here, not eloquence.
- 3
Correct the high weight sources
Review platforms and category directories will update on request. Roundup authors will usually correct a factual error if you email them. This is unglamorous outreach work and it is the highest yield activity in the whole discipline.
- 4
Give the web something specific to repeat
A number, a benchmark, a method with a name. Vague companies get vague summaries. Publish one defensible claim that a writer can quote and you will start seeing it come back at you in answers.
Share of model as the metric that replaces brand tracking
Brand tracking studies were always expensive, slow and a little bit theatre: a survey panel asked to recall names, run twice a year, reported with a wide confidence interval nobody in the room fully trusted. Share of model does the same job for a fraction of the cost and on a schedule you control. Run a fixed set of category questions across the major assistants, count how often you are named, correctly, and in the right context. That rate, tracked monthly, tells you whether the market's mental model of you is moving in the direction you want.
It will not be a smooth line. Assistants update retrieval and re-ranking behaviour without warning, and a correction to a single high weight source can move the number more than a quarter of content work. That volatility is a feature, not a flaw. It is telling you exactly which sources carry weight, which is more than a biannual brand survey ever did.
3
prompts is enough for a usable baseline: what, who, versus whom
4-5
external sources usually account for most of a wrong description
6wk
typical turnaround after correcting a stale review profile
Brand mentions in AI answers are not the same as brand awareness
It is tempting to treat any mention as a win, the way an old school PR team treated any coverage as good coverage. It is not the same signal. A brand mentioned in an AI answer as a cautionary comparison, or lumped into a generic list of five vendors with no distinguishing detail, is not building recall. It is filling space. The mention worth tracking is the one where the model attributes a specific claim to you: the fastest onboarding, the only vendor with a named integration, the lowest price in a stated range. That is the version of a mention a buyer actually remembers.
This is why counting raw mention volume across assistants is a weaker metric than it looks. Ten generic mentions a month is worse than two specific ones, because the specific ones are the only kind that change what a buyer believes about you before they ever open your site.
The counter-argument worth taking seriously
What to do when the model has you wrong
The uncomfortable case is not absence. It is a confident wrong answer: an assistant describing a product you retired, a price band you never charged, or a category you deliberately left. Absence is a content problem. A wrong answer is a correction problem, and it moves on a different clock.
- 1
Find the source, not the answer
Ask the assistant where the claim came from, then ask again in a fresh session. The same two or three URLs will keep appearing. Those pages are the actual target, not the assistant.
- 2
Fix your own surface first
If the wrong claim is defensible from anything you publish, including an old pricing page or a stale comparison table, fix that before asking anyone else to change theirs. Half the corrections I have chased ended here.
- 3
Send a factual correction, not a pitch
One paragraph, the wrong claim quoted, the correct claim, and a link that evidences it. No positioning language. Editors action these because a factual error is their problem too.
- 4
Re-check on a fixed cadence
Assistants do not re-read the web on your schedule. Check monthly for a quarter. Corrections that landed in third party sources usually show up in answers within one to two months.
Track corrections the way you would track a bug queue: what was wrong, which source carried it, when the request went out, and when the answer changed. After two quarters that log tells you which sources actually feed the models in your category, which is the map everyone else in your market is guessing at.
"We do not control third party sites, so this is not really actionable"
A reasonable objection is that this entire framework depends on getting other people's websites to change, and most of those sites owe you nothing. A review platform has no incentive to prioritise your correction request over the hundred others in its queue. This can look like an excuse to do nothing while waiting on someone else's schedule.
The lack of control is real, but it is not the same as having no way in. Review platforms and directories are built to keep listings accurate because inaccurate listings hurt their own credibility, so a specific, well evidenced correction request usually gets actioned within weeks, not never. Roundup authors respond to factual corrections far more often than to pitches, because a factual error is a small professional embarrassment they would rather fix quietly. The realistic expectation is not full control. It is a response rate high enough that four or five corrections a quarter meaningfully change the picture, which the worked example above shows in practice.
The strategic implication is uncomfortable for anyone who treats positioning as an internal exercise. A positioning statement that has not propagated to the sources a model trusts is not positioning. It is an internal document, agreed on in a room, that the market has not yet been told about.
What this changes about how positioning work gets scoped
Most positioning projects end at the website. A workshop produces a new narrative, a writer turns it into homepage copy, a designer builds the new hero section, and the project is marked complete. Under this model that sequence is maybe sixty percent of the job. The other forty percent is the unglamorous list: which review platforms carry a stale category, which roundups need a correction email, which forum threads from two years ago still describe the old product to anyone who searches for it.
Scoping that work up front changes the shape of a rebrand budget. It also changes who does the work. A copywriter cannot fix a stale G2 category tag. That task sits closer to lifecycle marketing or partnerships, and most teams have never assigned it to anyone because it did not used to matter enough to bother.
A short audit you can run this week
Baseline the current description before you touch anything
- ✓Ask three assistants what your company does, who it serves, and who it competes with. Save the raw answers.
- ✓List every named source in those answers and rate each one accurate, stale or wrong.
- ✓Rank the wrong and stale sources by how often they were cited across your prompt set.
- ✓Draft one canonical two sentence description and check it against every profile you actually control.
- ✓Set a date thirty days out to re-run the same prompts and compare.
None of this requires new tooling or a platform subscription. It requires someone to run the prompts, read the answers, and send the correction emails. The teams that treat this as beneath a strategy function are the ones still wondering why their AI answers describe a company they stopped being two years ago.
What I would do Monday
- 1Ask three assistants what your company does and screenshot the answers before you change anything.
- 2List every third party page named in those answers and check each description against your actual positioning.
- 3Write one canonical two sentence description and push it to every profile, listing and partner page you control.
- 4Pick the two most quoted category roundups in your space and get corrected or included.
- 5Re-run the same prompts in thirty days and log whether the description moved.
Common questions
- How do LLMs learn about your brand?
- Through a mix of what the model saw during training and what it retrieves live when answering a question. Both draw on the same pool of public pages: your site, review platforms, listicles, forums, press coverage. The model weighs sources it treats as credible more heavily than your own marketing copy, so a stale or wrong description on a widely cited third party page can outrank an accurate one on your homepage.
- What is share of model and why does it matter?
- Share of model is the rate at which an assistant names your brand, correctly and in the right context, across a fixed set of buyer questions. It matters because it is the closest available equivalent to search visibility for a channel that produces no reliable session data. A brand with strong rankings and weak share of model is invisible at the exact point buyers are forming a shortlist.
- Can you fix how an AI describes your company without a big rebrand?
- Usually yes, and faster than a rebrand would. Most wrong descriptions trace back to four or five external sources: a stale review profile, an outdated roundup, a forum thread nobody corrected. Fixing those sources, plus publishing one consistent description everywhere you control, moves the generated answer faster than new homepage copy does, because the model still has to trust the correction before it repeats it.
- Does having more content about your brand improve AI visibility?
- Not on its own. Volume without a specific, checkable claim gets treated as more of the same and gets summarised rather than quoted. A single sourced, dated statement that a model can lift whole will outperform a dozen generic pages saying the same vague thing in different words.
- How often should I check what assistants say about my company?
- Monthly at minimum, using the same fixed set of prompts each time so the comparison is apples to apples. Positioning changes propagate slowly through third party sources, so a weekly check will mostly show noise. A monthly cadence is frequent enough to catch drift and infrequent enough to be worth someone's actual time.
Who wrote this
Avishai Sam Bitton
Founder, DemandBox
Avishai runs demand generation programs for B2B SaaS companies across performance marketing, SEO, and answer engine optimization. He works directly with the teams he advises, with no account managers in between.
Connect on LinkedInThe long version
How to Get Your Brand Cited by AI Search Engines
A playbook for earning citations in AI answers: how retrieval works, before and after passage rewrites, the page checklist, and how to track citation share.
Want this argument applied to your numbers?
No deck, no discovery sequence. Tell us what you are spending and where the pipeline stalls, and we will tell you what we would change first.
Talk to an Expert