The method

Read. Choose. Prove. Spread.

Four moves, in an order that is not negotiable. Most of what goes wrong in this category is a team doing the fourth one first.

Taras NetrevFounder, ModelVoxUpdated August 2026

A claim you have not proved is a claim nobody will repeat. A claim nobody repeats is invisible to a machine.

Move 01

Read — find out what is being said.

We run the questions your buyers actually ask — not brand searches, but the category, problem and comparison questions that produce a shortlist — across ChatGPT, Claude, Gemini, Perplexity and Google AI Overviews. We record what comes back verbatim: whether you appear, what is said about you, who is named instead, and in what order.

Then the part almost nobody does: we trace each description to its sources. A model's characterisation of a company is not invented; it is assembled. Knowing which three pages produced the sentence is what makes the sentence changeable.

In parallel we check whether your site is legible at all — whether the content exists in the served HTML, whether the entity graph is coherent, whether crawler policy says what you think it says. This is where we most often find that a company has been invisible for a year for a reason nobody had looked for.

01 The Default Sentence reportWhat each major system says about you, verbatim, for the questions that produce shortlists — with the competitors it names alongside you.
02 Source traceThe specific pages and publications shaping each description, ranked by apparent influence. This is the map of what to change.
03 Legibility checkWhether your content survives a crawler that does not run JavaScript, whether the entity graph resolves, and what your crawler policy actually permits.
04 Competitive readThe same collection run against the two companies you are most often confused with, side by side.
The artefact

You leave this stage with a document that contains sentences a machine wrote about your company, next to the sources that produced them. In our experience it is the single most persuasive internal document a marketing team can hold, because it is not an opinion.

Move 02

Choose — decide the sentence that should replace it.

This is the positioning decision, and it is where most of the difficulty lives. The output is not a tagline. It is a claim about what category you are in, who you are for, and the one thing that is true of you and not of the companies you are compared with.

We hold every candidate claim to three tests, and they are unforgiving. Is it true — could a customer be produced who would confirm it under questioning. Is it specific — does it contain a noun a competitor cannot also use. Is it exclusive — is it false when applied to your two closest competitors. A claim that survives all three is rare, and it is worth the argument it takes to find.

Sometimes the answer is a new category. More often it is winning an existing one on better terms, because a category you cannot credibly lead is a liability rather than an asset — you spend years explaining a word instead of selling a product.

01 The claimOne sentence, tested against the three criteria, written to be repeated rather than admired.
02 Category decisionWhether to define a new category or reframe the existing one, argued both ways with the trade-offs made explicit.
03 Message architectureThe hierarchy from the single claim down to proof points, so that everyone in the company says a compatible thing.
04 Evidence ledgerEvery claim you make, mapped against where it is currently corroborated. The empty rows are the work.
05 Competitive differentiationA clear-eyed map of who you are genuinely compared with, and where the exclusive ground is.
Move 03

Prove — make the claim checkable.

A claim becomes durable when someone outside your company can verify it without asking you. That means producing the things that make verification possible: real numbers, published method, named customers, original research, a technical detail a competitor cannot match.

This is also where the site gets rebuilt, because a claim that a machine cannot read has not been made. Content in the served HTML rather than assembled in the browser. An entity graph that connects the company, its people, its services and its published work. Crawler policy that matches your intent rather than your host's default.

And occasionally this stage produces the answer nobody wants: the claim is not yet true. When that happens we say so. The work then becomes making it true — which is a product and operations conversation, not a marketing one, and it is a better use of the year than amplifying something the market will decline to repeat.

01 Evidence assetsThe specific artefacts that make the claim checkable: original research, published benchmarks, method documentation, named references.
02 Machine-readable siteServer-rendered content, coherent entity markup, clean internal linking, and a crawler policy that reflects a decision rather than a default.
03 Proof gapsA written list of the claims not yet supportable, and what would have to exist for them to be.
04 Publication planWhat to publish, in what order, so that the evidence exists before it is needed.
Corroboration distance

Our measure for how established a claim is: the number of independent, model-visible sources that repeat it. Zero is an assertion. One is a press release. Three or more is what gets restated to a buyer. Every claim in the ledger carries a number, and the programme is the work of raising them.

Move 04

Spread — get it corroborated where machines look.

The final move is the one that decides the outcome, and it happens almost entirely off your own website. Models weight what independent sources say. So the programme is aimed at getting the chosen claim repeated, accurately, by publications, analysts, communities, comparison sites, partners, customers and directories that the systems already read.

Some of that is conventional: coverage, contributed writing, analyst briefings, getting listed where a category is catalogued. Some of it is not: publishing research other people want to cite, so the corroboration arrives without being asked for. The second kind compounds and the first does not.

What we will not do is buy it. Purchased mentions, syndicated filler and engagement farms are visible to anyone paying attention — increasingly including the systems themselves — and a corroboration graph built out of them is a liability with a delay on it.

01 Corroboration mapThe specific sources that shape answers in your category, ranked by influence and by how reachable they actually are.
02 Owned researchOriginal work worth citing, designed so that the citation carries your claim with it.
03 Third-party programmeSequenced outreach to the publications, analysts, communities and catalogues that matter, with the claim held constant across all of them.
04 Re-read and reportThe Read stage, run again on a schedule, against the same questions — so movement is measured rather than asserted.
The frameworks

Five ideas we use, defined plainly so you can use them without us.

01

The default sentence

What a model says about a company when nothing specific about it can be corroborated: the category description with the company's name attached. Accurate, interchangeable, and read by your buyers.

02

Corroboration distance

The number of independent, model-visible sources that repeat a given claim. Zero is an assertion; three or more is what gets restated to a buyer. The number, not the wording, is what moves an answer.

03

The evidence ledger

Every claim a company makes, listed against where it is corroborated. Most ledgers are mostly empty on first pass. The empty rows are the roadmap.

04

The recommendation gap

A market with real demand where no company yet owns the recommended answer. The cheapest growth available, and the reason market choice belongs upstream of positioning.

05

Machine-readable positioning

The discipline itself: stating a position so that a system with no relationship to you, no interest in you and no follow-up questions can restate it accurately to a buyer.

Use them

These are yours

Definitions are not a moat and pretending otherwise is how a category fails to form. Use them, argue with them, cite them. We would rather the vocabulary spread than stay ours.

Frequently asked

On the method.

The questions we get most often once someone has read this far.

Read takes days: the collection runs in under a week, and the finding is a document, not a dashboard. Choose takes two to four weeks, most of it argument rather than production. Prove depends entirely on whether the claim is already true — if it is, weeks; if it is not, the honest answer is that the engagement changes shape. Spread runs in months and does not stop, which is why we scope it as a programme rather than a deliverable.
Four things. Whether you appear at all for the questions your buyers actually ask. What the description says when you do. How many independent sources corroborate the claim you have chosen — corroboration distance. And whether the technical layer is intact: does the page render without JavaScript, is the entity graph coherent, does crawler policy match intent. We report the ones that did not move as prominently as the ones that did, and we do not claim causation we cannot demonstrate.
Because it is the only stage that is expensive to reverse. Getting a claim repeated by independent sources takes months and a great deal of goodwill. Doing that for the wrong claim spends both, and then spends them again undoing it. The order is the discipline: nothing gets spread until it has been chosen and proved.
Both, deliberately. The site work — rendering, structure, entity markup, the pages that carry the claim — is ours because a claim handed over as a slide reliably dies in a backlog. What we do not do is content production at volume, media buying, or link acquisition, because none of those solve the problem and the first two make it worse.
Then the finding says so and the engagement should be small or none. That has happened. A company with a genuinely distinctive claim, a well-rendered site and a decade of independent coverage does not need us; it needs to keep going. We would rather say that in week one than sell twelve months of work to prove it.
Related answers

Questions buyers ask a model before they ask a firm.

How do AI systems decide which companies to recommend?
What is documented, what is strong inference, and what nobody actually knows.
Read the answer →
What is an AI visibility audit?
What one should contain, what it cannot tell you, and how to judge a vendor selling one.
Read the answer →
Which AI crawlers should you allow?
Every major AI crawler, what it is for, and a recommended policy — sourced from the operators’ own documentation.
Read the answer →

Start with Read.

It is the cheapest stage, the fastest, and the one that determines whether the rest is worth doing. The first call is complimentary and the finding is yours.

The first call is complimentary — and the finding is yours to keep.