Accounting Practices in France: Getting Recommended by AI Search

A Paris accounting practice with three partners and forty accountants can lose a generation of prospects for one reason: the information it publishes is not shaped like an answer. AI assistants do not rank firms; they assemble answers from passages they can parse, quote, and attribute. A page holding only a phrase is unusable to them. A page holding a claim that survives being lifted out of context is quotable.

The answer-object approach

The unit of AI-visible content is what we can call an answer object: a short, self-contained statement that resolves one question a real client would ask, with enough context to be trusted and enough specificity to be quoted.

Most firms already have the raw material — fee structures in proposal templates, onboarding steps in a PDF, recurring client answers in years of email threads. What they lack is the discipline of publishing one question per answer object. "Tax advisory" is a label. "A Paris practice handles corporate tax filings for small French limited companies, the deadline falls in the second quarter for calendar-year businesses, and the engagement is quoted as a fixed fee by turnover band" is an answer object: quotable, attributable, checkable against its source.

Four properties make a passage reliably quotable:

1. Self-containment. The sentence makes sense when lifted out of the page: no pronouns pointing back to a heading, no "as mentioned above", no reliance on an image to carry the meaning.
2. Specificity. Names the entity, jurisdiction, conditions, numbers, and timeframe. Vague claims get hedged or dropped; concrete ones survive.
3. Attribution. The page must make clear who is speaking. For a multi-partner firm this is not automatic.
4. Stability. The content must not contradict itself across pages, because a model finding two conflicting statements about an entity will usually refuse to use either.

That fourth point is where most practices quietly lose. The homepage says the firm specialises in "startups and scale-ups"; the services page says "traditional SMEs, liberal professions, and farmers"; the about page lists three offices in three towns. An assistant reading all three concludes the source is unreliable and moves to a competitor whose pages agree.

Building answer objects from what you already know

Take the twenty questions your team answers most often, and write each in the wording a client would use — "Do I need to audit my accounts?" rather than "Statutory audit". Then write one paragraph per question, structured the same way every time: direct answer first, conditions and exceptions after, contact path last. That repeatability is what makes a passage easy for a retrieval system to lift into an answer.

Entity clarity for a multi-partner firm

A multi-partner firm has a specific risk in generative search: it is described inconsistently by third parties and therefore looks like several different entities. Assistants key on the practice name, the city, the partner names, and the service list. If a model finds a partner's surname on a directory, the same surname on an article about regional tax practice, and the firm name on the firm's own site, it must decide whether these describe the same organisation — and ambiguity means the firm gets skipped.

Three things fix most of it:

• One canonical description, repeated verbatim. Decide the sentence that identifies the firm — full name, legal form, city, partner names, services listed — and reuse the same wording on the homepage, about page, contact page, every directory listing you control, and every professional profile you own. Not paraphrases. The same string.
• Partner-level pages. Give each partner a short page with a stable URL, a photograph, a role, and two or three sentences on what they personally handle. Partners are named entities in a way firms are not, and pages like these are frequently retrieved as the source for a person's expertise.
• Local specificity. Place names are unambiguous anchors that assistants weight heavily. Say which cities and departments you actually cover: "Clients in Lyon, Villeurbanne, and the surrounding Rhône department" is more retrievable than "southeastern France".

This is bookkeeping for the retrieval layer, and the cheapest work available to a small practice.

The monthly prompt panel

Structured content does not guarantee presence. Models change, sources rise and fall, competitors publish continuously. A practice that treats AI visibility as a one-off project is measuring nothing.

The alternative is a prompt panel: a fixed list of questions, asked on a fixed schedule, with the answers stored. Ten to fifteen prompts is enough. Build the list from your existing client questions and keep two or three control prompts you do not expect to win — broad questions where any credible firm might be named. Those show whether you are visible at all, while the specific prompts show what you are cited for.

The workflow is deliberately boring:

1. Run every prompt in a fresh session, and record the full answer text, the sources cited, and the date.
2. Score each answer for whether the firm is mentioned, whether the firm's own page is cited, and whether the description is accurate.
3. Diff the panel monthly, and classify each change as mention drift, description drift, or citation drift.
4. Investigate only changes that repeat. A single answer that shifted in one direction on one prompt is usually model noise.

That last point saves more time than anything else here. Assistants rewrite their phrasing constantly, so without a written record of the previous answer you will chase changes that were never real.

Bilingual and cross-border client questions

A meaningful share of the questions a French practice receives arrives in English, or in French with embedded English terms, from founders who did not grow up in France, international subsidiaries, and diaspora clients. If the firm's pages are French-only, those questions get answered without it.

You do not need a full English site. You need English answers to the English questions, on pages structurally identical to the French ones so the entity description does not fork. A realistic French query for a practice's core offering looks like « cabinet d'expertise comptable pour PME » — and the same client base searches close variations of it in English, usually with more urgency and less context.

The tricky part is not translation. It is that the two language versions must not describe the firm differently. Different fee philosophy, partner emphasis, or service lists fork the entity even when the translation is flawless.

What a small practice should measure

Traffic is not the metric — assistant answers often arrive without a click — and rank is not one either, because assistants do not produce a ranked list.

• Citation rate. The share of panel prompts where your own page is cited. This moves when your content changes, and it belongs in front of the partners.
• Prompt coverage. The share of your priority question set where the firm is mentioned at all, cited or not. Broad visibility across more questions is a stronger position than a high citation rate on a handful.
• Description accuracy. How often the assistant describes the firm correctly — which services, which city, which partner. Accuracy problems are invisible on a citation-rate dashboard and are frequently why a firm is mentioned but not recommended.
• Competitor displacement. How often a firm that appeared in your panel three months ago no longer does.

Keep the panel modest enough that someone will maintain it. Twelve prompts maintained monthly beats a hundred maintained once a quarter.

Summary

AI assistants do not recommend the best firm. They recommend the firm whose pages contain the clearest, most consistent, most quotable statements about what that firm actually does.

For a French accounting practice, that comes down to four habits: write one answer object per real client question and publish it on its own page; make the firm's identity identical across every surface you control; ask a fixed panel of questions monthly and keep the answers in writing; and treat description drift, not traffic, as the metric that matters. It requires the discipline to write answers instead of brochures.

Frequently Asked Questions

How long does it take for an AI assistant to start citing an accounting practice? No dependable timeline exists, and anyone offering one is guessing. Observe instead: does the firm appear at all in answers to its own priority questions, and do those answers cite the firm's own pages rather than a directory listing? Structured answer-object pages are often picked up within weeks because they are easy to parse and attribute. Ranked visibility for competitive questions takes far longer and is never guaranteed. Track citation rate rather than waiting for a win you cannot schedule.

Should a small French practice publish in both French and English? Only if it genuinely serves clients who work in English, and even then only for pages where the language changes the answer — cross-border tax questions, international filings, English-language client onboarding. A mirrored English site usually splits your authority and creates maintenance drift, and a page existing in two languages with slightly different claims can confuse readers and assistants alike. Publish English where demand is demonstrable, keep the entity description identical across languages, and leave the rest in French.

How do we stop an assistant describing our firm incorrectly? Start by finding where the wrong description comes from. Usually it is a third-party directory, a professional register entry, or an outdated practice profile someone else controls, rather than your own pages. Fix the sources you own first, keeping the wording of your canonical description identical everywhere, then correct the third parties through their own update channels. While those changes propagate, publish the facts on your own site — founding partners, city, service list, legal form.