Paris can enter an AI answer even when nobody invited it. The lab examines whether this happens because the capital is genuinely relevant, or because source density makes Paris the easiest version of France for a model to describe.
A category prompt can look innocent: “best independent repair workshops in France,” “French bakery examples,” “agencies for a regional trade category.” In one composite run, the lab gave the model a city frame outside Paris, then widened the prompt. Lyon held for a while. Marseille appeared as a broad contrast. Smaller cities thinned quickly. Paris returned like a default setting, not always named first, but often carrying the densest descriptions and the most confident wording.
The odd part was not that Paris appeared. For many categories, it should. The odd part was how easily the answer borrowed Paris-shaped evidence when the query asked for a wider French frame. A business outside the capital might be described through a directory snippet. A Paris example received a fuller paragraph, sometimes with press-like texture. One answer even kept a non-Paris business name but described its category in language that sounded copied from a capital-centred overview. The lab marked the case as a source question, not a complaint about geography.
Defining Paris skew without turning it into a slogan
Paris skew — in this material — is the tendency of an AI answer to make Paris-based or Paris-framed evidence more visible than other French city evidence, because that source layer is easier for the model to retrieve, connect or justify. The definition matters because the lab is not saying every mention of Paris is an error.
Some categories are genuinely concentrated in the capital. Some national organisations have Paris headquarters. Some trade bodies, press coverage and English summaries naturally point there. The question is narrower and more useful: when a user asks about a business category across French cities, does the model preserve the city frame, or does it drift toward the source layer with the most repeated and readable evidence?
The lab reads this through recorded observations: prompt wording, query language, city frame, answer text, visible citations, missing sources and recognizable source-like traces. A conclusion comes only after several related runs. One answer that starts with Paris is not enough. Several answers that weaken Lyon, Marseille, Lille, Rennes, Strasbourg or smaller-city examples while thickening Paris descriptions become material for a pattern.
This is not a ranking study. Sourceplane Atelier does not claim measured shares of Paris visibility. It compares source dependencies. The more precise question is whether the answer is city-faithful or source-convenient.
How the comparison is built
The lab uses a simple comparison structure. It asks category questions with explicit city frames, then repeats them with regional and national frames. It also changes query language. A French prompt may retrieve local business directories, municipal language, trade pages or French press. An English prompt may retrieve summaries that explain the category to outsiders, and those summaries often lean toward better-known cities.
The composite Lyon bakery from the research plan helps here because it is familiar enough to be visible and local enough to be fragile. When the prompt asks directly about Lyon, the answer may keep the city, the category and a local listing. When the prompt asks for examples of French bakeries or business visibility in France, the same kind of business can become less specific. Paris examples may receive fuller context because more sources are available and easier to paraphrase.
The composite regional repair network between Brittany and Provence exposes a different version of the skew. Instead of bakeries and local food, the category involves branches, parent identity and uneven press mentions. The model may give Paris-based or nationally framed examples a cleaner explanation, while regional branches are described as fragments. Sometimes the answer does not place Paris first. It simply gives the capital-style evidence more weight.
That distinction is important. Skew can live inside the paragraph, not only in the ordering of names. A smaller-city business may appear but be described thinly. A Paris business may be treated as the category prototype. The user leaves with a sense of hierarchy even if the list looks balanced at first glance.
The source layers that make Paris easier
Paris has a source-density advantage in many business categories. More national press mentions, more English-language explainers, more headquarters pages, more travel and service summaries, and more copied lists make capital evidence easier for a model to retrieve. The lab does not need to romanticize this. The machinery prefers evidence it can connect.
A directory-led answer may still include non-Paris businesses, especially when the prompt is local. But directories often supply thin fields: name, address, category, rating-like traces, hours, sometimes a short description. If Paris examples are supported by richer press or company pages, the final answer can feel uneven even when the model did not explicitly decide that Paris is better.
A registry-anchored answer can reduce some skew by grounding businesses in official identity, but registry evidence has its own narrowness. It can confirm that an establishment exists and where it is registered. It does not supply reputation, category nuance or local meaning by itself. A smaller-city company may be registry-visible and still commercially under-described.
Press-amplified answers are where Paris skew often becomes warmest. A press mention gives the model a story, not just a record. If Paris businesses have more narrative coverage, they can be described with more authority and color. Regional press can counter this, but only when it is retrievable and connected to the business identity. Otherwise, local specificity remains trapped in sources the answer does not use.
Region-flattened answers complete the pattern. The model names the category, maybe names a city, but loses the department, branch relation or regional business context. Once those details disappear, Paris does not even have to dominate openly. The rest of France becomes a softened background.
When Paris is relevant and when it is lazy evidence
The lab avoids the easy reaction: remove Paris from the answer and call it fair. That would be bad research. Paris is often relevant. It may be the administrative centre, the strongest press node, the place where a trade body sits, or a legitimate comparison point for a category.
The problem begins when Paris becomes the proxy for France without being named as a proxy. A model may answer a national question using capital-centred examples because they are better documented. That is understandable. But if the answer then sounds like it describes the market evenly, the reader receives a distorted map. The distortion is smooth, not crude.
A useful answer would make the source condition visible. It might say that Paris examples are easier to document, while regional examples require city-specific sources. Most AI answers do not pause to make that distinction. They present a coherent paragraph. The lab has learned to distrust coherence when the source trail is uneven.
For SMB readers, this is especially relevant outside the capital. Weak visibility does not always mean weak business identity. It may mean the public source layer is thinner, less connected, more local, less English-readable or less often repeated. A regional business can be well known in its market and still be difficult for an AI system to justify in a broad answer.
What agencies and trade bodies can learn
Agencies often ask whether their clients need more content. The lab’s answer is usually less tidy. More content can help if it clarifies city, category, branch relation and source consistency. More generic content may simply add another soft paragraph to the pile. The issue is whether the business has evidence that a model can connect without sanding off the place.
Trade bodies face a similar tension. A national page may list members or categories in a way that helps AI systems recognize the sector, but it can also strengthen the national frame over regional distinctiveness. If a trade body wants Breton, Alsatian, Provençal or smaller-city members to remain visible as local businesses, the page structure has to preserve those details. Region is not a caption. It is part of identity evidence.
The lab also notes a language problem. English-language summaries of French business categories often explain France through the most internationally familiar places. That can make Paris an interpretive shortcut. French-language prompts may recover more local records, but they can still lean on national directories or broad category pages. Testing both languages is therefore necessary.
The most practical diagnostic is a city swap. Ask the same category question with Paris, Lyon, Marseille and one smaller city. Then ask it nationally. Watch what survives: the source types, the business names, the branch distinctions, the descriptive richness. If the smaller city appears only as a name while Paris receives the explanation, the skew is already visible.
Limits of the finding
This material does not measure the size of Paris skew across all French AI answers. The lab does not present a sample percentage, a city ranking or a fixed rate of distortion. Its method is comparative and descriptive: repeated prompts, source-layer reading, language shifts and city frames that another reviewer can reconstruct.
There are cases where Paris dominance is not distortion. Luxury categories, national administration, finance, media, cultural institutions and some B2B services may legitimately produce more Paris evidence. A careful reading must ask whether the category itself is capital-heavy before treating the answer as skewed. The lab’s concern is with unexplained drift, not legitimate concentration.
Visible citations do not settle the matter either. A cited Paris source may be only one part of the answer path. A non-Paris source may be used without citation. A directory can include smaller-city records while the generated description still borrows its tone from broader summaries. The lab therefore separates cited dependency from probable dependency and avoids pretending that a source card reveals the whole route.
The conclusion is restrained but useful. Paris skew is best read as a source-layer effect. AI systems often favor evidence that is denser, easier to retrieve, more repeated or easier to justify. If regional business evidence remains thinner or less connected, the answer may keep France in the title while letting Paris write the paragraph. Whether that changes depends on source conditions, and the lab marks that as an uncertainty note rather than a promise.