AI Mentioned You. That Is Not a Recommendation.

An injection-moulding supplier sent me a screenshot last month, pleased with himself. He had asked an AI assistant where to source small-batch precision parts, and his company appeared in the second paragraph. He read it as a recommendation.

I asked him to read the sentence out loud. It said: "suppliers capable of this work include A, B and C among others."

He stopped halfway through.

Appearing in an answer and being recommended are two different outcomes, and most companies never check which one they got. That single distinction decides whether the rest of your effort is pointed in the right direction.

This article does not offer a way into a recommendation list, and it cannot promise one. It addresses a narrower question: of the appearances you can see, which kind are you looking at.

Three positions, three different values

Recommended. The wording runs "recommend", "worth considering", "a good fit for". The company sits in a candidate position, usually with a condition attached - suitable for this volume, experienced with this material.

Mentioned. The wording runs "including", "such as", "among others". This is enumeration. The system knows the company exists and makes no claim about fitness. The supplier above landed here.

Excluded. The wording runs "be careful with", "avoid suppliers that". This is a cautionary reference, and it is worse than silence, because it arrives attached to a reason a buyer will remember.

The three are not degrees of the same thing. They are different outcomes produced by different questions, and they call for different responses.

The wording is the tell

You do not need a tool to tell them apart. Read the sentence around your name.

• "Recommend", "suitable", "preferred for" - candidate position.
• "Including", "such as", "among others" - enumeration only.
• "Be cautious", "avoid", "watch out for" - negative reference.

A hundred listings that all read "among others" amount to a hundred instances of the system knowing you exist. That is not nothing, but it is not what most people think they have.

The harder case is the third one. Companies are frequently used as a cautionary example without knowing it, because nobody screenshots an answer that criticises them.

Why the position shifts with the question

The same company can sit in all three positions within one afternoon, depending on how the question is phrased.

A recommendation question - "which suppliers would you suggest for 500-piece runs" - forces the system to produce candidates. Only here does a recommendation position exist.

An information question - "which companies in this region do precision machining" - requires only a list. Being included satisfies the request; no judgement of relative fitness is produced.

An exclusion question - "what should I watch out for when choosing a supplier" - invites a cautionary example, and the example may be a category rather than a company.

This is why "we showed up in AI" is not a useful metric on its own. It does not say which question produced the appearance, and the question determines what the appearance is worth.

Platform structure adds a second variable. Different engines draw on different shelves of sources, so the same question returns different company sets depending on where it is asked. Measuring one assistant and generalising to all of them is the most common analytical error in this work.

From listed to recommended, one thing is missing

Enumeration requires existence. Recommendation requires a reason.

For a system to place you in a candidate position, there has to be a sentence it can produce: what you are better suited to, which conditions you handle, what can be verified about your capability. That sentence has to come from your published material. It is not invented.

This is the part most content programmes skip. They produce volume - more pages, more posts, more coverage - which improves the odds of being enumerated. Volume does not create reasons. A thousand pages saying "quality and service" still yields nothing a system can use to distinguish you from the company next door.

So the shift in focus is not from invisible to visible. It is from visible to defensible.

Why volume does not convert into recommendation

The instinct when you discover you are only being listed is to publish more. More pages, more posts, more coverage. That instinct is half right and mostly wasted.

Volume raises the probability of being enumerated, because there are more surfaces on which the system can encounter your name. It does nothing for the probability of being recommended, because recommendation requires a reason and reasons are not produced by repetition.

A thousand pages saying "quality, service, reliability" produce a thousand identical claims. To a retrieval system comparing twelve suppliers, that is one signal repeated a thousand times, not a distinguishing one. It cannot use it to place you above anyone, because it does not separate you from anyone.

The analogy that tends to land with sales teams: a candidate who sends a thousand identical CVs gets read more often than one who sends three. He does not get shortlisted more often, because the CV still does not say what job he fits.

What changes the position is a claim a machine can compare. Not a bigger claim - a comparable one. The next section is about what that looks like in practice.

What a defensible claim actually looks like

The difference between a listed company and a recommended one is usually a handful of sentences, and those sentences share a property: they can be checked.

Instead of "high precision", name the band and the boundary. "Plus or minus 0.01 mm on turned parts up to 40 mm, plus or minus 0.05 mm above that" is a claim with an edge. It tells a system what you handle and, implicitly, what you do not.

Instead of "fast delivery", name the range and the variable. "18 days standard, 25 days in the fourth quarter" is checkable. "Fast" is not, and a system comparing four suppliers with "fast" on all four pages has no basis for ranking them.

Instead of "experienced", name the duration and the shape of the operation. "Operating since 2009, 60 staff, two shifts, in-house tooling" gives a reader something to weigh. "Rich experience" gives nothing.

Instead of "customers worldwide", name the sector and the use case. Naming sectors rather than clients avoids the confidentiality problem while still telling a system which questions you are relevant to.

Instead of "certified", name the standard and the revision. "ISO 9001:2015" is a fact; "certified quality management" is a mood. Revisions matter, because the version number is what makes the claim verifiable at a point in time.

None of these requires disclosing anything commercially sensitive. Each converts a paragraph of positioning into a sentence another system can hold up against a competitor's sentence. That is the entire mechanism by which a listing becomes a recommendation.

One test worth applying to every page: cover the company name and read the paragraph. If the paragraph still reads as true of any competitor, it is not a reason - it is a placeholder occupying a page.

How to audit where you actually stand

Most companies believe they know their position and are working from one screenshot taken six months ago. A short audit replaces that with something usable.

Pick three recommendation questions in the language your buyers use - not your product category, the way a buyer phrases the request for help.

Ask each of them in every assistant that matters to your market. Record the position: candidate, enumerated, absent, or negative. Record the surrounding wording verbatim, because the wording is the diagnosis.

Repeat the same set two weeks later. Generative answers vary between sessions. A single reading tells you one possible outcome; two readings tell you whether it is stable.

Write the result as a one-page table - question, assistant, position, wording. This page is more useful than any ranking report, because it shows which question you are losing on rather than an aggregate score that hides it.

The output of the audit is not a score. It is a short list of questions where you are enumerated but not recommended, which is precisely the list of reasons you have not yet built.

Five things that build a reason

State who it is for. Not "we serve a wide range of clients", but the specific case your capability fits: this material, this volume band, this tolerance, this certification regime.

State the difference as a fact, not a verdict. "Two-week standard lead time where the regional norm is four" is a fact. "Better than competitors" is a verdict, and verdicts about third parties tend to cost more than they earn.

State what you are not suited to. This is the least intuitive item on the list and the one that most reliably improves how a page reads. A supplier who says "not economical below 200 pieces" is more credible on everything else, because the boundary is specific enough to be checked.

State verifiable quantities. Lead time in days, minimum order in pieces, tolerance in millimetres, pass rate as a percentage. Adjectives are not quantities. Certifications belong here too, with the exact standard and version named.

Keep the vocabulary identical across platforms. If the website says "small-batch" and the marketplace listing says "flexible production", a system sees two capabilities. Worse, when sources disagree, retrieval tends to fall back to the most conservative version of the claim.

A workable test: hand the five items to someone outside the company and ask them to repeat your positioning in one sentence. If they cannot, the reasons are not yet reasons - they are still adjectives arranged in paragraphs.

Three boundaries worth stating

Recommendation cannot be bought. Google's published policies are explicit that paid placement does not influence generative results, and that content is not guaranteed to be surfaced. What you can influence is whether a defensible reason exists.

Position is not stable. Generative answers are produced per query, per session. The same question asked twice in a week can return different sets. Read trends, not single screenshots.

Different assistants disagree. Because they read different shelves, the same company can be a candidate in one and absent in another. That is normal, not a malfunction, and it means coverage has to be checked per engine.

Closing

Being mentioned means your material entered the corpus. Being recommended means the system holds a reason to place you among candidates.

The first is an entry ticket. The second is a seat.

An assistant knowing you exist is not the same as an assistant vouching for you.

Three questions that come up

If my name appears in the answer, is that already a recommendation?

Not necessarily. Read the surrounding wording. "Recommend" and "suitable for" indicate a candidate position; "including" and "among others" indicate enumeration. Most appearances reported internally are enumeration.

What do I do if I am being used as a cautionary example?

Find the specific claim behind it first - usually one figure, one certification, or one lead time that is contested or outdated. Correct that claim at the source, then publish material that addresses the same concern directly. Negative references almost always trace back to a single disputed fact rather than to a general reputation.

Do different assistants give different positions?

Frequently. Each draws on a different set of sources, so the same question produces different company sets. Check each engine separately before drawing a conclusion.

How long does it take to move from listed to recommended?

There is no reliable answer, and anyone who gives you a number is guessing. Two variables dominate: how quickly the engine rebuilds its index, and how widely the new claims get republished across third-party sources. Indexing can be a matter of days; being picked up and echoed by other sites takes longer and is not under your control. What you can control is whether the claim exists and is consistent, which is the precondition for any of it.

Do we need different content for each assistant?

No. The underlying claims should be identical everywhere - that consistency is what makes them usable. What differs is coverage: each engine reads a different set of sources, so a claim published in one place may reach one assistant and not another. Write the claims once, publish them consistently, then measure each engine separately rather than assuming a single placement covers the field.

The short version

Three positions exist: recommended, mentioned, excluded. The wording around your name tells you which you are in. Moving from mentioned to recommended requires a reason a machine can state, and reasons are built from specifics - audience, difference, boundary, quantity, consistent vocabulary - not from volume.

Stop counting appearances. Start reading what they say.