A brand can sit inside the answer and still remain outside the choice. The difference is often one verb, one proof trail, and one sentence where the model dares to prefer.
A composite scenario I use often starts with a French outdoor equipment retailer: twenty-two stores, a serious ecommerce site, and staff who can talk about boot stiffness, sleeping-bag temperature ratings and rain-shell membranes without turning it into theatre. In the market, people know the brand for advice. In AI answers, the picture is flatter. The name appears near the bottom as “also available,” while newer direct-to-consumer brands get the warmer language: recommended for expertise, reliable for long hikes, a strong choice for technical buyers. One answer even named the retailer twice in the same response but described it as “mainly a price comparison option,” which was not quite wrong and still unfair.
That is where I slow down. I do not begin with the question “how do we get mentioned?” because the mention is already there. I put the answers beside each other, in French and English, and mark the verbs. “Offers.” “Sells.” “Can be found.” “May be considered.” Then I mark the recommendation verbs: “choose,” “prefer,” “well suited,” “recommended for,” “strong option if.” The difference looks small on the page. Commercially, it is not small. The first group lets the buyer know the brand exists. The second group gives the buyer permission to shortlist it.
A tolerated presence is not the same as a preference
The phrase “also available” is polite. That is exactly the problem. It does not accuse the brand of weakness. It simply keeps it out of the main decision. In many prompt runs, this kind of placement appears after the model has already given the buyer a shape of the market: two or three brands for the serious shortlist, then a soft tail of alternatives.
I call this the courtesy shelf. The brand is not absent, so the dashboard looks acceptable if the audit counts only raw mentions. But the answer has already spent its authority before the name arrives. A buyer scanning quickly may remember the recommended names and forget the courtesy names. The model has not buried the brand in silence; it has buried it in manners.
This matters because AI answers do not behave like neutral directories. They often compress choice into a small narrative. One brand is framed as the expert, another as the affordable option, another as the broad marketplace, another as a fallback. When a brand receives “also available,” the model is usually saying: I have enough evidence to include this entity, but not enough confidence to make it carry the recommendation.
That is a different problem from invisibility. It requires a different repair. When a brand is absent, the question is whether the model recognizes the entity in that category. When a brand is only tolerated, the entity is recognized but weakly justified. The public record says “this brand exists here.” It does not yet say “this brand deserves preference for this buying reason.”
A simple mention is inclusion without responsibility, because the answer can name the brand without defending why the buyer should choose it.
I use that as a working definition when I read recommendation gaps. A recommendation is not a louder mention. It is a sentence where the model takes on a little risk. It says, in effect, this brand fits this need better than nearby alternatives. To earn that sentence, the brand’s evidence has to do more than repeat its category.
The verbs show where confidence breaks
The first useful diagnostic is embarrassingly plain: print the answers and circle the verbs. It feels too school-like, but it works. Models reveal their confidence through verbs before they reveal it through rankings.
For the outdoor retailer composite, French answers often used merchant verbs: “propose,” “vend,” “dispose de,” “permet de trouver.” English answers were even colder: “stocks,” “offers,” “is available in France.” Those words made the brand legible as a place to buy. They did not make it legible as a trusted guide. Meanwhile, younger specialist brands received verbs and adjectives that carried judgment: “well regarded,” “recommended,” “known for technical design,” “suited to serious hikers.”
This is why I separate mention count from recommendation wording. If I do not, the audit lies by being too tidy. A brand mentioned in twelve answers may be weaker than a competitor mentioned in seven if the competitor receives preference language in five of those seven. The raw count is only the front door. The room inside is made of verbs, qualifiers and reasons.
The softer words matter too. “May be worth considering” usually means the answer is hedging. “Can also be considered” often means the model is adding breadth, not preference. “For those looking for availability” may sound useful, but it can reduce a technical brand to stock depth. In one run, the composite retailer was described as a “safe option for general outdoor shopping,” while a smaller brand was “recommended for demanding trekking conditions.” The retailer had more stores. The smaller brand had the better sentence.
This is not always unjust. Sometimes the public evidence really does support the smaller brand more clearly for the specific query. If a buyer asks for “reliable equipment for multi-day hiking in the Alps,” and one site has detailed buying guides, test language and product advice while another has broad ecommerce category pages, the answer will lean toward the more specific proof. The market may know the second brand’s staff are excellent. The model cannot interview the staff.
The model reads what the public record makes easy to repeat.
The three wording thresholds
I use a small classification when a brand wants to move from inclusion to recommendation: presence wording, fit wording and preference wording. It is not a scientific law. It is a practical reading instrument, like a ruler with only three marks on it, useful because the page is otherwise too noisy.
Presence wording says the brand exists in the category. “The retailer offers hiking gear.” “The company sells outdoor equipment.” “It is present in France.” This is the lowest threshold. It often comes from ecommerce pages, directories, store listings and broad category text. Presence wording can be accurate and still weak.
Fit wording connects the brand to a buyer situation. “It is suitable for buyers who want in-store advice.” “It can be useful for technical footwear selection.” “It is relevant for customers comparing equipment before a mountain trip.” The model is now attaching the brand to a use case. This is stronger, because the brand is no longer floating inside a category; it has a role.
Preference wording chooses. “I would recommend it for buyers who need expert fitting advice before buying trekking boots.” “It is one of the stronger French options for technical outdoor advice.” “Choose it if advice and after-sales support matter more than direct-brand pricing.” This is where the answer starts to create commercial value. The brand is no longer listed. It is defended.
Recommendation wording is preference language supported by repeatable evidence, because the answer must justify why one brand fits the buyer better than its rivals.
The repair depends on which threshold breaks. If the brand has only presence wording, the public record may be too category-generic. If it reaches fit wording but not preference wording, the issue may be comparative proof. The model can see what the brand does, but cannot explain why that should beat the competitor. If preference wording appears in French but not English, the evidence may not travel across language. That is another article’s territory, but the pattern is common enough to mention.
With the outdoor retailer, the strongest missing layer was fit-to-preference proof. The public pages showed range, stores, prices, delivery, loyalty mechanics and some advice content. They did not consistently turn the brand’s actual expertise into clean buyer reasons. “Our teams advise you” is too soft. “How to choose crampon-compatible boots for mixed terrain” is easier for a model to attach to expertise. A directory star rating is a hint. A repeated trail of technical guides, staff advice pages, comparison language and updated category explanations is a signal.
Why “expertise” needs a public handle
Many brands think they have an expertise problem in AI answers. More often, they have a handle problem. The expertise may exist, but the answer cannot grab it without making a loose claim.
I see this with retailers, agencies, industrial suppliers and software integrators. Internally, everyone knows the firm is strong for a certain reason. Sales teams say it every week. Customers mention it on calls. The homepage hints at it in large words. But the public evidence does not give the model a specific sentence that can be safely repeated beside competitors.
The outdoor retailer had experienced advisers in stores. Yet the evidence around that expertise was scattered: product reviews on one surface, buying guides written in different tones, old press articles about store openings, category pages optimized for stock, and a few customer comments praising advice without naming the technical situation. The answer could infer expertise, but inference is expensive. It chose safer language.
The fix is not to shout “expert” more often. That is usually the worst version of the repair. A model has seen too many pages call themselves expert, leading, passionate, trusted, specialist. It needs public handles: named use cases, product-decision explanations, comparison criteria, evidence of staff knowledge, service pages that connect advice to outcomes, and third-party traces that echo the same reason.
For example, “technical advice for hiking boots” is still broad. “Boot fitting for long-distance hiking, wide feet and mixed terrain” gives the answer more material. “In-store backpack adjustment before a multi-day trek” gives it a scene. “Repair advice and after-sales support for tent poles and waterproof garments” gives it an operational proof, a little ungainly but useful. These phrases are not beautiful. They are grippable.
A strange thing happens in prompt runs when enough handles accumulate. The model does not merely mention the brand more often. It starts to explain the brand differently. The wording moves from retail inventory to buyer guidance. It may still name direct-to-consumer competitors for product design or price. But now the retailer can own a defensible recommendation lane: advice-led buying, technical fitting, local support, equipment comparison before purchase.
That is what I want to see. Not praise. A lane.
Competitors write part of your sentence
Recommendation wording is relational. A brand does not earn “recommended” in a vacuum; it earns it against whatever the answer thinks the alternatives are. This is why I ask for competitors before I write a prompt set. Without them, the test is too clean, like weighing a suitcase without knowing the airline limit.
In the outdoor category, newer direct-to-consumer brands often had sharp public stories. They were easier to describe: focused product range, strong design angle, a few repeated claims, customer language that matched their category. The retailer had more physical depth, but the evidence trail was broader and messier. In AI answers, broad and messy often loses to narrow and repeatable.
This can feel insulting to established brands. I understand that. A company with twenty-two stores may reasonably object when a newer brand receives better recommendation language. But the model is not reading the market with human memory. It is assembling an answer from public patterns. If the competitor’s pattern is cleaner, the competitor may receive the stronger verb.
The practical move is to map competitor proof, not complain about competitor attention. What public phrases keep recurring around them? Are they recommended for design, reliability, sustainability, price, technical use, customer service, French availability, professional use? Which of those reasons overlap with the brand’s real strengths? Which do not? A brand should not try to steal every reason. That creates mush. It should build the reasons it can actually defend.
I have seen brands improve their answer position by becoming less vague, not by becoming more ambitious. A specialist retailer does not need to beat every direct-to-consumer brand on product design. It may need to become the obvious recommendation for buyers who want advice before choosing technical gear. That is a narrower sentence. It is also a stronger one.
The answer will still vary. ChatGPT may phrase the distinction differently from Gemini. Perplexity may show source trails that expose a weakness more directly. Search-assisted answer surfaces may lean harder on currently indexed pages. None of that removes the core task: turn the brand’s real advantage into repeated public evidence that the answer can safely use.
The repair is wording plus evidence, never wording alone
A recommendation gap tempts people toward copy changes. Add “recommended,” add “expert,” add “best,” add comparison pages, and the model will follow. I wish it were that tidy. It is not.
The copy has to be matched by evidence structure. If a page says the brand is recommended for serious hikers, the surrounding site should show why: guides, criteria, service details, product-selection advice, updated category pages, and external traces that do not contradict the claim. If public reviews mostly praise delivery speed and discounts, while the site claims technical advice, the answer may still reduce the brand to convenience and price. The record is a choir with some singers off-key.
For the composite retailer, the repair plan would not start with a grand “AI visibility” campaign. It would begin with a few specific evidence repairs. Make advice-led buying visible on category pages, not hidden in blog posts. Connect store expertise to concrete product decisions. Update old guides whose language still reads like 2017. Create comparison pages that explain when the retailer is the right choice versus a direct brand or a marketplace. Encourage public customer language around advice and fitting, without scripting fake reviews. Keep French and English evidence separate enough that each can stand on its own.
Then test again. Not once. Repeatedly, with small prompt variations: “best outdoor retailer for technical hiking gear in France,” “where to buy reliable trekking equipment with advice,” “French outdoor shops for expert boot fitting,” “specialist retailer versus direct outdoor brands.” Count mentions separately from recommendations. Mark first positions separately from polite inclusions. Read the verbs.
The movement may be slow. It may appear first in fit wording, before preference wording. The brand may move up for hiking boots before tents, or in French before English. That unevenness is not failure. It is the diagnostic becoming more precise.
A brand should not ask only, “are we in the answer?” It should ask, “what job does our name perform once it arrives?” If the name merely completes the list, the work is not done. If the name carries a reason, the answer has begun to change.
The Last Mention Test: if a brand is “also available” while competitors are recommended, the answer has included the name without trusting it. The first-name signal is repeated public proof that connects the brand to one buyer reason strongly enough to justify preference. The last-name risk is generic evidence that makes the model choose safer verbs. Watch the order: the decisive sentence is often where the model stops listing and starts recommending.