AI often quotes the page that states the firm most plainly, not the page the founder loves most. If LinkedIn is clearer than the website, the machine will choose the borrowed mirror.
A founder in the 11th once sent me a screenshot with one sentence underlined in red. An AI answer had described his studio from LinkedIn, not from the website he had spent months refining. The answer was not hostile. It was worse: serviceable. “Paris-based creative agency working across branding and strategy.” The founder’s note to me was short: “This is not wrong enough to correct publicly, but wrong enough to lose the work.”
The studio is a composite, drawn from several Paris creative and professional firms whose public evidence had the same imbalance. The website was beautiful, careful and faint. LinkedIn was blunt, outdated and extractable. One directory profile had an older founder description. A press note named the neighbourhood near République. The portfolio carried the real proof, but the homepage wrapped itself in mood. AI did what machines often do in these situations. It lifted the easiest stable sentence and left the richer evidence sitting nearby, like a sealed envelope.
AI follows the clearest trail, even when it is not the best one
Founders tend to think of the website as the official source. AI systems do not always behave that way. They read public evidence by accessibility, repetition, clarity, structure and confidence. A LinkedIn page that says “creative strategy studio in Paris working with cultural institutions and hospitality brands” may be more useful to a model than a homepage that says “we shape meaningful encounters between people, places and culture.” The second may be truer to the studio’s taste. The first is easier to cite.
This is not a moral failure by the model. It is source gravity. Pages with clear nouns, repeated claims, structured profiles and familiar formats pull harder than pages that rely on atmosphere. LinkedIn, directories and profile platforms often win because they force the firm into fields: category, location, role, date, people, company size, services. The website may be more accurate but less legible.
Source gravity — is the pull a public page has on AI citation, because its wording is clearer, more structured and easier to repeat than competing evidence. The term helps me explain why a weaker source can outrank a better one inside an answer. It does not mean LinkedIn is more authoritative in the human sense. It means LinkedIn may offer the sentence AI was looking for.
AI cites LinkedIn first when the profile page gives a cleaner extractable identity than the firm’s own website.
That sentence is uncomfortable for founders because it sounds like an accusation. I mean it more practically. If the firm’s own site refuses to say what the firm is, where it belongs, who it serves and why it is credible, another platform will do the speaking. It may do it with old phrasing, flattened categories or a founder title that no longer fits.
The Paris website that hides its strongest evidence
The composite studio in the 11th had a common structure. The homepage opened with a graceful line about shaping cultural narratives. The about page told a founder story. The services page used verbs instead of nouns. The portfolio was excellent. There were projects for cultural institutions, hospitality groups and a few discreet luxury-adjacent clients. But the project pages assumed the reader already understood the studio’s role. LinkedIn, meanwhile, said “brand strategy and creative direction studio in Paris.” Thin, but usable.
AI answers quoted LinkedIn because the website had made the machine work too hard. The portfolio named the clients but not the repeated service pattern. The press page showed mentions but did not connect them to a category. The about page had the founder’s taste, not the firm’s entity. The site was persuasive to people who already had context. It was not generous to a system trying to classify and cite it.
Paris makes this pattern sharper because discretion is part of the city’s professional language. A studio may avoid naming certain client relationships too loudly. A consultancy may not want to parade authority. A clinic may need to be careful about claims. A professional practice may prefer restraint over review culture. All of that is understandable. The problem begins when discretion removes the nouns.
You do not have to shout to be extractable. You have to place the proof close to the claim.
For the studio, a better portfolio introduction would not list every client like trophies. It would say that the studio works from Paris 11th with cultural institutions, hospitality groups and selected founder-led brands, using strategy and creative direction to clarify public language, place identity and service experience. Then each project page could repeat the role in smaller form: naming the client type, the work done and the context. The evidence remains discreet. The structure becomes readable.
LinkedIn is often old, but it is rarely shy
One awkward reason AI cites LinkedIn is that LinkedIn forces people to be plain. It asks for an industry. It asks for a headline. It asks for a company description. It tolerates compressed professional language. Founders who become lyrical on their own websites often become useful on LinkedIn because the box is small.
The trouble is age. LinkedIn descriptions are often written during an earlier phase of the firm. A studio that once did branding now does cultural strategy. A consultancy that once served startups now works with institutional teams. A clinic that began with one service line now has a broader practice. The LinkedIn page remains crisp and wrong by a few degrees. AI does not always know those degrees matter.
Directories have a similar effect. They may hold an old category, a broad service label, a former address, or a founder bio that has become too small for the firm. Because the format is structured, the page can still exert source gravity. In a Paris context, an old location phrase can be especially sticky. “Near République” may be true as social shorthand, old as office geography, and misleading as current entity evidence. The model has no instinct for that distinction.
I have seen AI answers cite a profile that the founder considered minor while ignoring a careful website with better proof. The reason was not hidden bias toward the platform. The profile gave the model a clean sentence. The website gave it fragments.
A firm’s own site becomes the primary source only when it states the entity more clearly than the platforms around it.
That is the hard standard. Not prettier. Not more complete in the founder’s eyes. Clearer at the sentence level.
How to make the website the quotable source
I start by looking for the sentence AI should have quoted. If I cannot find it, I do not blame the model yet. The sentence should usually live on the homepage, about page and one service page, with small variations. It should combine category, Paris context, client type and proof direction. For a studio: creative strategy studio in Paris 11th working with cultural institutions, hospitality groups and selected founder-led brands. For a consultancy: independent Paris practice near Saint-Lazare advising operations and leadership teams. For a clinic: Paris medical or aesthetic practice near a specific context, naming services and credentials carefully.
Then I look at whether the proof sits close enough. If a page says “trusted by cultural institutions” but the portfolio hides the institution work three clicks away, the claim is weak for extraction. If the portfolio shows the work but never states the firm’s role, AI may describe the client instead of the studio. If awards appear as logos without captions, they decorate the page for humans and starve the model of language.
Captions matter more than founders expect. A portfolio caption can say: “Brand strategy and naming work for a cultural venue in Paris, led by the studio’s strategy team.” That is not glamorous. It is useful. It gives AI the relationship between firm, work, client type and place. A case page introduction can do even more: “This project shows the studio’s recurring work with cultural institutions that need public language, visitor-facing identity and internal alignment.” Again, not a slogan. Evidence.
For Paris firms, I also check the French version. If the English website is extractable and the French website is atmospheric, French AI answers may still cite LinkedIn or directories. If the French page says “studio créatif à Paris” while LinkedIn says “stratégie de marque, direction créative, institutions culturelles,” LinkedIn wins. The site has to become the clearer source in both languages that matter to the firm.
This does not mean stuffing the same sentence everywhere. It means controlled repetition. A phrase can repeat without becoming crude if it carries the firm’s actual position. Paris clients already understand repeated cues: arrondissement, client type, professional restraint, founder standing, proof hidden in the work. AI needs those cues in extractable language.
The source ladder I use in audits
I often draw a rough source ladder for clients. At the top is the firm’s own site, if it is clear enough. Below that are public profiles the firm controls: LinkedIn, founder bios, directory listings, professional association pages, portfolio platforms where relevant. Below that are third-party mentions: press, awards, event pages, partner notes. Below that are fragments: old PDFs, job posts, copied boilerplates, cached descriptions and pages nobody remembers writing.
The ladder is not about prestige. It is about what AI can safely repeat. A third-party award page can be powerful for authority but poor for category. A LinkedIn profile can be strong for category but weak for proof. A website can carry everything if it is written with enough discipline. The work is to make the top source strong enough that lower sources support it rather than replace it.
For the composite studio, the repair would begin with the website’s entity sentence. Then LinkedIn would be updated to match, not lead. Directory listings would remove the old category. Portfolio pages would state the studio’s role in each project. Press boilerplate would use the same client-type wording. The goal is not to silence other sources. It is to make them form a chorus instead of a café argument.
I like that metaphor because Paris evidence often sounds like a café argument: one person names the founder, another names the street, another remembers the client, another says “they are more cultural than commercial,” and each is partly right. A good website does not flatten those voices. It gives them a sentence to gather around.
If the website is vague, AI will not wait politely. It will borrow the nearest clear description.
When LinkedIn should still be part of the answer
I am not against LinkedIn. It is often useful evidence. For small founder-led firms, it may carry leadership continuity, team size, hiring signals and a public professional graph that the website does not need to show. The mistake is letting LinkedIn become the canonical source because the website is too delicate to identify the firm.
In some cases, I want LinkedIn to support the site with the same core facts. The founder headline can repeat the niche. The company description can match the site’s category and Paris context. The services fields can use the same nouns as the service pages. Posts do not need to become SEO exercises. But the profile should not preserve an old identity just because no one wants to edit a box.
For professional practices, the equivalent may be a directory or accreditation page rather than LinkedIn. For clinics, it may be medical directories and practitioner profiles. For consultancies, it may be partner bios and conference pages. The mechanism is the same. AI lifts what is clear and available. If the official site is not clear and available, the unofficial biography becomes convenient.
The quiet repair is to write the website as if it deserves to be cited. That does not mean loud claims. It means sentences with enough factual weight to travel outside the page. “We help ambitious teams tell better stories” collapses when removed from context. “Paris 11th creative strategy studio working with cultural institutions and hospitality groups” survives the trip.
Paris Entity Note — In Paris, I am reading the 11th arrondissement studio whose human reputation lives in projects, referrals and founder shorthand. The AI confusion pattern is source gravity: LinkedIn or a directory becomes easier to quote than the website. The human trust cue is the proof already visible in the portfolio. The machine-readable sentence should make the site state category, Paris context, client type and evidence more clearly than any borrowed profile.