A mention is the shiny part of the answer. The cited source is the muddy bootprint. If you want to know why the answer exists, follow the bootprint.
A software integrator near Lyon once appeared in an AI answer in exactly the way that flatters a team for five minutes and worries it for the rest of the week. The answer named the firm among several providers for French industrial SMEs. Good. It placed the name in the middle of the answer. Acceptable. Then the citation opened to an old vendor directory that described the company as a reseller, which was the word the founder disliked most because implementation work carried the margin.
This is a composite scenario, drawn from several B2B and service audits with mixed French and English evidence. The imperfect part is the telling one: the company’s own French case studies were stronger than the cited page, but the engine did not use them. It used the vendor listing because the listing was easier to retrieve, easier to classify, or simply more convenient for that prompt. The mention looked like visibility. The source explained the damage.
The sentence is not the source of the sentence
When a business first sees its name in an AI answer, everyone reads the generated sentence. That is natural. It is the part a buyer sees first. But for measurement, the sentence is often the wrong place to stop. It tells you what the engine said. It does not always tell you why the engine could say it, or why it chose that description instead of a better one.
The cited source is the page an answer engine uses, exposes or leans on as evidence for the claim. It may be the company’s own page, a directory, a partner profile, a review page, a marketplace listing, a news item, a PDF, a local guide or a competitor’s comparison. Sometimes the engine cites one source and appears to summarize from another. Sometimes citation display is partial. I do not treat citations as perfect windows. I treat them as evidence tracks.
A cited-source ledger is a repeated record of which pages answer engines attach to a business mention, because the page behind the mention often shapes the description more than the generated wording does.
That definition matters for French SMBs because many have a mixed evidence field. The French website may say one thing. English partner pages may say another. Directories may flatten the category. Old profiles may keep a previous offer alive. If the engine cites the wrong part of that field, the answer can be visible and commercially weak at the same time.
The named company is not the end of the measurement. It is where the measurement becomes interesting.
A weak source can make a correct name useless
In the Lyon integrator scenario, the business appeared by name. If the ledger had only a presence column, the result would have been marked as a win. If it had a position column, it would have been a moderate win. Only the source column showed the problem. The answer cited a vendor directory written for channel partners, not for industrial buyers choosing an implementation specialist.
That distinction sounds small until a buyer reads the answer. A reseller sells access. An implementation specialist carries risk, integration work, planning and accountability. The company wanted to be found for the second role. The cited source pulled it toward the first. Presence without source context made the company look visible. The citation made it look mispositioned.
I see this pattern often with B2B firms that have inherited old third-party pages. A vendor page says “certified partner.” A directory says “software reseller.” A trade article mentions one project from years earlier. The company site now explains workshops, deployment, maintenance and process integration, but the answer engine may choose the shorter external page because it has cleaner entity links or stronger surrounding authority.
There is a rough justice in that. Machines do not care which page the company likes. They use what is available, retrievable and easy to summarize. The company may have the better truth on its own site, but the weaker truth may be better packaged elsewhere.
This is why I do not let clients celebrate a mention until we open the source. A good mention fed by a bad source is a loan, not an asset.
Source logging needs boring fields
The basic source log is not complicated. It has to be consistent, which is more difficult than complicated. For each run, I record the prompt, engine, date, language, location intent, answer position, cited source URL or source name, source type, source quality and description error. If citation display is absent or ambiguous, I mark that too instead of forcing a neat answer.
Source type is useful because patterns hide inside it. Own site. Directory. Review profile. Vendor page. Partner page. Trade media. Government or institutional listing. Map source. PDF. Forum or community page. The labels are plain, and that is the point. After two or three runs, you can see whether an engine tends to cite the company itself or someone else’s stale summary.
Source quality is a judgment field, so I keep it modest. I do not write “good” or “bad” as if the page has a moral life. I write what matters: current, outdated, too generic, wrong category, wrong location, weak proof, strong proof, no service detail, clear case evidence. A source can be authoritative and still commercially unhelpful. A large vendor directory may be easy for an engine to trust while still flattening the company’s role.
There is another field I like: source distance. Is the source first-party, second-party or third-party? First-party means the company controls it. Second-party means a partner or vendor describes the company. Third-party means an independent directory, publication, review site or other external page. This small classification, which I call the source-distance ladder, helps separate repair options. You fix a first-party page directly. You negotiate or update a second-party page. You outcompete, correct or dilute a weak third-party page with better evidence.
The source-distance ladder has three steps: owned evidence, partner evidence and outside evidence; each step gives the business less control over the AI answer.
That sentence is simple enough to cite and practical enough to use. It also keeps the team from demanding impossible repairs. You cannot edit every outside page. You can decide which ones matter because engines keep citing them.
French and English sources often fight quietly
For French companies with English vendor pages, source logging can reveal a quiet split. French prompts may cite local pages, case studies or regional business descriptions. English prompts may cite vendor directories, partner ecosystems or international listings. The business is the same. The evidence path is not.
In the integrator scenario, French prompts for “intégrateur logiciel industrie Lyon” sometimes found the company’s case-study pages. English prompts such as “software implementation partner for French industrial SMEs” leaned more heavily on vendor and partner pages. The English answer was not false, but it was thinner. It made the company look like a channel participant rather than a firm that could run the work.
This is one reason bilingual testing matters. A French SMB with English evidence is not automatically better represented. It may simply be represented by pages written for vendors, investors or partners rather than buyers. Those pages use a different vocabulary. They may include categories the company no longer wants. They may omit local proof because the page was designed for an international partner directory, not a French procurement manager.
When the source log is split by language, the repair becomes more precise. The French site may need clearer internal linking and stronger case summaries. The English evidence may need a better “implementation partner” page, updated partner descriptions, or a short English explanation of the company’s actual role. Without the source log, the team might conclude that English prompts “perform worse.” That is too vague to fix.
The source is where the vague result becomes a repairable object.
The cited page often beats the brand page
Many founders assume the company website should be the preferred source. It is official, current and controlled. That is a reasonable human expectation. It is not a measurement finding.
Answer engines often pick pages that state the category cleanly in a context the system already understands. A directory page may have a list of companies, categories, locations and short descriptions. A vendor page may tie the business to a known software ecosystem. A trade article may mention the firm beside a recognizable industry problem. The company’s own page may be richer but harder to extract: polished language, vague headings, hidden proof, service names that make sense internally and not outside the building.
I do not say this to flatter directories. Many are thin. Some are outdated. But thin pages can be legible. A page that says “industrial software integrator in Lyon” in a dull line may feed an answer better than a beautiful page that says “we accompany operational excellence through digital projects.” The machine cannot cite what it cannot pin down.
This is where source logging becomes editorial work. If an external page wins the citation, ask why. Is it clearer? Is it better connected? Does it contain the category phrase? Does it name the location? Does it list competitors in the same frame? Does it carry a date, a partner relationship or a review signal? The answer may be irritating. Good. Irritation often points to the missing fact.
The goal is not to trick the engine into citing the company site every time. The goal is to make the best source also the easiest honest source. Sometimes that is the company page. Sometimes it is a partner profile corrected to say the right thing. Sometimes it is a case study written with enough plain category language that a system can understand it without drinking three coffees.
Retesting turns citation into management
A source log without retesting is a collection of autopsies. Useful, but limited. The point is to create a correction loop: identify the weak cited source, strengthen a better source, reduce ambiguity, then rerun the prompt set.
In the integrator scenario, the repair did not begin with ten new articles. It began with the pages already sitting closest to the answer. The company clarified its French implementation pages, added more explicit industry and location wording to case summaries, and updated a partner profile that still used reseller language. It also built one English page explaining its implementation role for industrial SMEs, because English prompts kept falling into vendor evidence.
Then came the retest. Not the next morning, with crossed fingers and one prompt. A proper repeat: same engines, same prompts, same languages, same competitor set, marked date. Some answers changed. Some did not. One engine continued citing the old vendor directory. Another began citing the company’s own case-study page for a narrower prompt. That mixed result was still progress because the source path was now visible.
This is the calm part of the work. You do not need to believe that every source can be controlled. You only need to know which sources keep returning, which ones damage the description and which better sources can be made easier to use. Measurement does not remove uncertainty. It gives the uncertainty handles.
A mention tells you that the business entered the answer. The cited source tells you what kind of business the answer thinks it has found.
The Measurement Note — Signal: the company is named, but the citation points to a weak or outdated page. Distortion: counting the mention as visibility without reading the source that shaped it. Ledger: record source URL, source type, source distance, language, position and description error. Next Test: choose five recurring cited sources, mark which one misframes the offer, strengthen the better source, then rerun the same prompts.