The machine can know your name and still misunderstand your work. That is the dangerous middle state: visible enough to be repeated, wrong enough to send the buyer elsewhere.
In a composite scenario from B2B services, a small industrial calibration and maintenance firm near Nantes appears in an AI answer about providers for French manufacturers. At first glance, the team should be pleased. The company is named. It sits in the answer with several larger service groups. There is even a citation, though the cited page is an old distributor listing rather than the firm’s current service material.
Then someone reads the description out loud. The answer says the company mainly sells measurement equipment. It barely mentions on-site calibration, preventive maintenance or the awkward factory-floor work that brings in the useful contracts. In one run, it also implies the firm serves national retail buyers, when its real work is regional and industrial. The name is right. The commercial meaning is bent. That is not a small copy error; it is a visibility failure with a polite face.
Presence can hide damage
The first field in many AI visibility reports is presence: did the business appear or not? Presence matters, but it is crude. It is like checking whether a parcel reached the building without checking whether it went to the right room, with the right contents, on the right day. A business can appear in an answer and still lose the buyer because the description points to the wrong offer.
I see this most often with companies whose public evidence is scattered across partner pages, old directories, case studies, event listings, review profiles and service pages written in different years. The model gathers a shape from those fragments. Sometimes the shape is close enough. Sometimes it preserves an old fact because that fact is easier to find than the current one. Sometimes the most citable source is the least accurate source.
For the Nantes calibration composite, the answer’s mistake is not random. Distributor listings often use broad labels: supplier, reseller, authorised partner, equipment provider. Those labels may be true in purchasing language and still weak in buyer language. If the company’s own pages do not clearly state calibration scope, maintenance work, industrial customer type and service area, the generated description may settle for the easier label.
Wrong AI description is a visibility failure because the business is present in the answer under an inaccurate commercial identity, which can misdirect buyers while making the team believe visibility has improved.
That definition is strict because the harm is strict. If the answer says you do the wrong thing, serve the wrong location or fit the wrong customer, the mention becomes contaminated evidence.
Accuracy deserves its own score
I score description accuracy separately from presence, citation and position. If those fields are blended, the report becomes mush. A high-position mention with a wrong description is not “mostly good.” It is a high-position error. A low-position mention with a precise description may be a weaker visibility signal but a stronger evidence signal. The distinction matters when deciding what to fix.
My usual accuracy scoring is plain enough to survive a busy meeting. Accurate. Mostly accurate with a small omission. Partly wrong. Wrong in a commercially meaningful way. Unclear or too vague to judge. I do not pretend this is laboratory science. It is a practical grading habit, and it forces people to read the answer instead of admiring the appearance of the brand name.
The important phrase is “commercially meaningful.” In a recurrent ledger pattern, an AI answer that gets a founding year wrong may annoy the founder, and it should be logged, but it may not change buyer behaviour. An answer that describes a maintenance specialist as a product seller can change the short list. An answer that says a regional firm serves the whole country can attract the wrong lead and repel the right one. An answer that calls a commercial-maintenance provider an emergency-only service may fill the inbox with low-margin requests.
This is where the ledger earns its keep. I use an accuracy field, an error type field and a short note. The note is not an essay. “Calls firm a distributor; omits calibration; cites old listing.” That is enough to start tracing.
Four kinds of description error
The errors repeat more than people expect. I use a classification called the four description breaks: offer break, location break, customer break and evidence-age break. It is not a grand theory. It is a sorting tray.
An offer break happens when the answer gets the service wrong or too narrow. A calibration firm becomes an equipment seller. A heating-maintenance firm becomes an emergency plumber. A strategic agency becomes a content producer. The company is visible, but under the wrong shelf label.
A location break happens when the answer stretches or shrinks the service area. Some AI answers love to tidy geography into broad regions. A business near Nantes becomes “France-wide,” or a regional operator disappears from nearby towns because the cited source only names one city. Location errors are especially costly for French SMBs because buyer intent often has a local hinge.
A customer break happens when the answer names the wrong buyer type. B2B becomes consumer. Industrial becomes general SME. Multi-location service becomes one branch. This error can be subtle. The words sound positive, yet the buyer reading the answer does not recognize their own situation.
An evidence-age break happens when the model repeats an old fact that still lives in a strong source. Former services, old partner tiers, outdated branch data, previous positioning, expired certifications. The company website may have moved on. The machine may not have followed, especially if the old source is clearer than the new one.
These four breaks help keep the conversation grounded. Instead of saying, “The AI got us wrong,” the team can say, “We have an offer break fed by a distributor listing” or “We have a location break in French prompts around Nantes.” That sentence already contains the beginning of the repair.
The cited source is often the source of the wrongness
When an answer misdescribes a business, people rush to edit the homepage. Sometimes that is useful. Often it is premature. First I want to know where the wrong description came from, or at least which cited source sits beside it.
In engines with visible citations, the work is direct. Open the cited page. Read the language. Does it contain the wrong label? Does it omit the current service? Does it use a generic category? Does it show an old address or old partner status? If yes, the answer may be less mysterious than it felt.
In engines with weaker citation display, the work is less clean. I look for repeated phrasing across public sources. If the answer calls the firm a distributor, I search the company’s public evidence for that label. Partner pages, directories and old announcements often carry the phrase. I also compare French and English prompts. A wrong English description may trace back to supplier language; a wrong French description may trace back to thin local pages or directories.
There is a rough annoyance here: the wrong source may be outside your direct control. A distributor listing, an old partner page, a trade-media profile, a public directory. That does not make the problem untouchable. It changes the correction loop. You strengthen the sources you control, request corrections where reasonable, and retest. If the better source becomes easier to cite and the old source less dominant, the description may improve across repeated runs.
I do not promise instant repair. Anyone who promises that has not watched enough answer ledgers. The work is iterative because the evidence pool is iterative.
Vague praise can be an accuracy problem too
Wrong descriptions are easy to spot when they contain a false fact. The harder case is soft blur. “A provider of business solutions.” “A trusted partner for technical projects.” “A company serving professional clients.” Nothing there is exactly false. It is also almost useless.
For B2B firms, vague praise can be commercially damaging because it erases the reason to choose the company. The Nantes calibration firm does not need to be called a provider of technical solutions. It needs the answer to understand on-site calibration, preventive maintenance, French industrial sites, service boundaries and the difference between selling equipment and keeping equipment reliable in production. If those details disappear, the buyer cannot judge fit.
I score vagueness as an accuracy issue when it hides the buying distinction. This is slightly subjective, so I write the criterion in the ledger. Does the answer state the core offer? Does it name the relevant customer type? Does it preserve location or market scope? Does it avoid a misleading label? If not, “vague” is not a harmless style note. It is a weak description.
The cure is not to make every page longer. Length can make fog thicker. The cure is to state the business facts in extractable language: what the company does, for whom, where, with what boundaries, and with which proof. Case studies can help if they are written as evidence, not as victory poems. Partner pages can help if they do not swallow the company’s own identity. Service pages can help if they say the plain thing before the elegant thing.
A model cannot repeat a clear fact that no public source bothers to say clearly.
Do not fix accuracy with flattery
When a team sees a wrong AI description, the first emotional reaction is often to add stronger claims. More authority. More adjectives. More ambition. That rarely helps. The answer engine does not need the company to sound larger. It needs the public evidence to become less ambiguous.
For the calibration composite, I would avoid language like “leading technical partner for industry.” It sounds impressive and says too little. I would rather see plain statements: calibration of specific equipment categories for French industrial sites; preventive maintenance where true; service area; distinction between equipment sales and maintenance work; examples that tie the work to real client situations without exposing private details.
The correction loop starts with the wrong fact. Then it asks which source might feed that fact. Then it strengthens or corrects the better source. Then it retests the same prompts. If the description improves once and fails three times, the loop continues. If it improves across engines and languages, the change becomes a measured repair rather than a hopeful edit.
This is also why screenshots are dangerous here. A screenshot can hide the error if the brand name is circled in red and the description is left unread. In a proper audit, I make the wrong sentence visible. It goes into the ledger. It gets a type. It gets a likely source. It gets retested. The mistake stops being embarrassing fog and becomes a repair ticket.
There is no shame in finding these errors. Most companies have old public traces. Most service descriptions were written for humans already inside the category, not for machines assembling evidence from fragments. The failure is not having a wrong description once. The failure is counting the mention as success while the wrong description keeps returning.
The better visibility question
A better question than “Do we appear?” is “When we appear, what are we?” It sounds almost philosophical, but in the ledger it is practical. Are you an emergency-only operator, a maintenance provider, a reseller, an implementation specialist, a local branch, a national network, a B2B firm, a consumer service? The answer’s category decision influences the buyer’s next click, even when the company name is correct.
I want French SMBs and marketers to become less excited by presence and more demanding about identity. Visibility is not simply a light switched on. It is a description repeated in public. If the description is wrong, more visibility can spread the wrongness faster.
So I keep the fields separate. Presence: yes or no. Position: where. Citation: which source. Accuracy: how correct. Error type: what broke. Next action: which source or page deserves attention. This is not complicated work, but it is fussy work. Fussy is good here. The alternative is a smooth report that misses the sentence doing the damage.
In the Nantes calibration scenario, a useful finding would not be “AI visibility is improving.” It would be more like: the company is present in supplier-category prompts, weaker in maintenance-intent prompts, often cited through old distributor evidence, and repeatedly described as an equipment seller rather than a calibration and maintenance specialist. That sentence hurts a little. It also tells the team where to work.
And that is the point. A wrong description is not an awkward detail after the real measurement. It is part of the measurement.
The Measurement Note — Signal: the business is named but described under the wrong commercial identity. Distortion: counting presence as success while the answer misstates offer, location or customer type. Ledger: record the wrong phrase, error type, cited source, prompt language and repeat count. Next Test: rerun the same five prompts after strengthening the clearest source, then score accuracy separately from presence.