A mention proves the system can find the brand. A recommendation proves the system can defend choosing it. The distance between those two sentences is where much of the repair work sits.
In a composite scenario from B2B software, a French integrator kept appearing in AI answers about retail and logistics systems. The name was present often enough that the marketing team felt half-relieved. It showed up in ChatGPT. It showed up in search-assisted answers. Perplexity surfaced a few sources that were genuinely relevant. Nobody could call the brand absent.
Yet the wording stayed weak. The answer would recommend two larger firms, describe another as “well suited for complex multi-site deployments,” and then add the integrator as a possible option. In one run, the brand appeared under “other companies to consider.” In another, the system named it but attached no reason at all. The rough detail was almost comic: one answer correctly identified its retail work, then called it a “consulting boutique,” which the company would never use and clients would not recognize.
Being found is the lower bar
A mention has a modest meaning. It says the model has enough evidence to connect the brand to the question. That evidence may come from the company’s site, directories, partner pages, articles, reviews, comparison pages or older fragments. The name enters the answer because the system can place it somewhere in the category.
A recommendation asks for more. It requires the answer to spend confidence. The system has to say, directly or indirectly, “this is a good choice for the buyer’s need.” That shift changes the grammar. The brand is no longer a loose example. It becomes a supported option.
The advice threshold is the line between being named and being chosen, because recommendation wording requires stronger comparative proof. I use that term when a brand appears frequently but never receives the verbs and adjectives of preference. The name is in the room. The answer does not hand it the microphone.
This distinction is uncomfortable because many dashboards count the mention and move on. A brand may celebrate visibility while losing the persuasive part of the answer. In a buyer’s reading, that difference is not theoretical. The recommended brands form the shortlist. The merely mentioned brands decorate the periphery.
Verbs show the status of the brand
When I read AI answers, I mark verbs before I mark conclusions. “Leads,” “specializes,” “is known for,” “is well suited,” “offers,” “may be considered,” “also provides.” These verbs carry rank. They reveal whether the answer is placing a brand in the center of the buyer’s decision or keeping it in reserve.
In French, the difference can be quiet but sharp. “Se distingue par” gives the brand a reason to stand out. “Est reconnue pour” offers inherited credibility. “Propose également” is weaker. “Peut convenir” is cautious. None of these phrases is fatal alone. Across repeated runs, they become a pattern.
The integrator in the composite case received many low-commitment verbs. It “offers integration services.” It “can support retail projects.” It “may be relevant for logistics companies.” These are not insults. They are thin endorsements. The larger competitors received stronger verbs: they “support complex digital programs,” “bring recognized expertise,” or “are often chosen for enterprise deployment.” Some of that wording came from the companies’ own public language. Some came from third-party descriptions. Some was the model smoothing the evidence into a confident sentence.
A mention without recommendation often means the brand has category evidence but lacks decision evidence. The model knows what the company does. It does not see enough public proof for why a buyer should choose it over the neighboring names.
Adjectives make the hesitation visible
Adjectives are small gauges. They show how much confidence the answer is willing to attach to the brand. A recommended brand is often “specialist,” “recognized,” “strong,” “experienced,” “reliable,” or “well suited.” A merely mentioned brand becomes “possible,” “alternative,” “smaller,” “regional,” “available,” or “also relevant.” The exact words vary, but the temperature changes.
This is where the answer can be unfair in a familiar way. A French company with deep client relationships may be softened because its public evidence is modest. A larger competitor may receive confident adjectives because its pages, partner materials and English summaries repeat the same claims. The model does not sit in sales meetings. It reads what the outside record gives it.
In the integrator’s notes, the adjective problem was tied to proof placement. The strongest case studies were buried in long PDFs. Some client names were public, but the use cases were not written in a way that matched buyer prompts. English pages were thinner than French pages, and the English wording leaned on general trust language. “Experienced team.” “Tailored support.” “Business solutions.” These phrases sound harmless. They do not teach the answer why the brand should be recommended for a specific buyer situation.
The competitor pages were not perfect. One used inflated language that a careful human would discount. Yet they attached adjectives to evidence more cleanly. “Retail deployment” sat near a named project. “Multi-site operations” sat near a case fragment. “ERP integration” appeared beside partner language. The answer could borrow the link.
Recommendation wording needs proof with a shape
A recommendation is built from proof that has shape. By shape I mean the evidence carries a clear relation between buyer need, brand capability and reason to prefer. A page that says “we support companies in their digital projects” has weak shape. A page that says “we integrate retail inventory systems across store networks and warehouse operations” has stronger shape. A third-party mention that confirms the same work gives the shape a second edge.
This is not a call for rigid copy. Human readers hate pages that sound assembled from keywords. The trick is to write plainly enough for the buyer and specifically enough for the machine. I prefer slightly rough clarity over elegant fog. A sentence can be imperfect and still be useful if it tells the answer where to place the brand.
For the integrator, the missing shape had three parts. First, the public pages did not repeat the company’s strongest category with enough consistency. Second, the evidence did not often compare the brand’s fit against larger generalist firms. Third, English-language proof did not carry the retail and logistics specialization strongly enough. As a result, the AI answer could name the company, but recommendation wording flowed toward competitors with cleaner public signals.
A useful repair starts by collecting the exact recommendation phrases competitors receive. Do they get “best for enterprise scale,” “strong for technical depth,” “good for mid-market implementation,” or “recognized specialist”? Then ask whether your public record gives the answer permission to say an equally specific sentence. If it does not, the gap is not only wording. It is evidence architecture.
The mention can be a useful starting line
I do not treat mentions as worthless. A brand that appears without recommendation is ahead of a brand that disappears completely. The system has found a connection. There is something to work with. The danger is stopping there.
A mention tells us which evidence is already visible. It may reveal that a partner page is doing more work than the homepage. It may show that a directory still carries an old category. It may expose a language split, where French answers know the brand and English answers barely know what to do with it. In that sense, the weak mention is diagnostic. It is a pale stain on paper that shows where the liquid entered.
The next step is to separate three counts: mentions, recommendations and first-position placements. For this article, the middle count matters most. How often does the brand receive recommendation wording? Under which prompts? In which language? Against which competitors? Which adjectives attach to it? Which verbs stay cautious? A brand that is mentioned eight times and recommended once has a different problem from a brand mentioned three times and recommended twice.
In the composite case, the integrator did receive a stronger recommendation when the prompt included “specialist retail logistics partner in western France.” That was revealing, though not sufficient. It meant the evidence could support a recommendation under a narrow phrasing. The repair was to widen the set of buyer questions where that proof became visible.
Repair the reason, then the sentence
Teams often want to rewrite pages so that AI will recommend them. That can become theatrical. The better order is to repair the reason first. Which buyer situation should the brand own? Which proof supports that ownership? Which sources repeat it? Which competitor currently receives the sentence that should partly belong to the brand? Which public fragments contradict or dilute the claim?
Once the reason is clear, sentence work matters. Service pages need direct category language. Case studies need summaries that a buyer and an AI answer can both understand. Partner descriptions should not hide the use case. English pages should not become a polite translation that loses the commercial edge. Third-party profiles should avoid dead labels that keep pulling the brand into the wrong bucket.
Recommendation wording does not arrive because a page asks for it. It arrives when the public record makes recommendation feel safe. That safety is comparative. The answer must see why the brand fits the buyer’s request beside the other names in the category. Without that, the model will keep the brand in a soft paragraph where nobody has to choose it.
The Last Mention Test: if the brand is visible but never recommended, the answer has found the name without trusting the choice. The first-name signal is proof that joins one buyer need to one defensible reason to prefer the brand. The last-name risk is category evidence with no decision edge. Watch the order: a mention opens the door, but recommendation wording decides who is invited to the table.