Skip to content
search-engine-optimizedby Netsolution

What assistants answer, and who they borrow it from

Twelve frozen buying questions, put periodically to conversational assistants. What they answer is recorded, which sources they cite, and whether their answer contradicts a published document. Here is the protocol — no record has been made yet.

State on :protocol frozen at version 1, no measurement carried out. This page will be completed by the records, in the order they arrive.

Why the protocol appears before the results

Because a protocol published after the first measurement cannot be told apart from a protocol written to justify its result. This one is frozen, dated and versioned in the site’s repository: any change to a question will read in the history, and nobody — the author included — will be able to claim it was always worded that way.

It is also this site’s first protocol. Since its very first page it has announced four claim statuses and has only ever used three: no claim carries tested, for want of a replayable experiment. This instrument is what was missing for the fourth status to stop being decorative — and it will carry that status only once it has produced results, not before.

The twelve questions

They cover the five directory categories and real buying decisions. 6 of them bear on a point settled by a published document: those are the ones that let a gap be recorded, and not merely a preference.

The questions are asked in French, word for word as printed below in French.The English wording is a reading translation and is never sent: a translated question is a different question, drawing on other sources and producing another answer. Both are shown so that nothing about the measurement has to be taken on trust.

  1. “Which tool should I use to track my site’s positions on Google?”

    Asked verbatim: « Quel outil pour suivre les positions de mon site sur Google ? »

    Which tools are named, in what order, and whether any source accompanies the answer.

  2. “Semrush or Ahrefs, which one should I choose?”

    Asked verbatim: « Semrush ou Ahrefs, lequel choisir ? »

    Whether the answer decides, on what criteria, and whether it mentions that entry tiers do not compare directly.

  3. “How can I tell whether a page of my site is indexed by Google?”

    Asked verbatim: « Comment savoir si une page de mon site est indexée par Google ? »

    Whether the engine’s console is cited first, or a third-party tool is offered for information no third-party tool holds.

    Reference document: Google Search Console, « About Search Console »

  4. “Is there a free alternative to Screaming Frog?”

    Asked verbatim: « Existe-t-il une alternative gratuite à Screaming Frog ? »

    Which open tools are named, and whether their vitality is checked — several commonly cited projects are abandoned.

  5. “How much does a Semrush subscription cost per month?”

    Asked verbatim: « Combien coûte un abonnement Semrush par mois ? »

    Whether an amount is given, whether it is dated, and whether it corresponds to monthly billing or to an annual plan reduced to a month.

    Reference document: Semrush, page tarifaire

  6. “Is the number of backlinks a ranking factor?”

    Asked verbatim: « Le nombre de backlinks est-il un facteur de classement ? »

    Whether the answer separates what the documentation says from a reading of the market, and whether it confuses number with quality.

    Reference document: Google Search Central, règles anti-spam de la recherche

  7. “Should I create an llms.txt file to be cited by AI?”

    Asked verbatim: « Faut-il créer un fichier llms.txt pour être cité par les IA ? »

    The engine writes that it does not use those files. An assistant recommending one contradicts a public document from its own publisher or from a direct competitor.

    Reference document: Google Search Central, « Google's Guide to Optimizing for Generative AI Features on Google Search »

  8. “Which backlink analysis tool suits a French-language site?”

    Asked verbatim: « Quel outil d'analyse de backlinks pour un site français ? »

    Whether French-speaking vendors are named, or whether the answer stops at the English-language suites.

  9. “What is crawl budget, and do I need to deal with it?”

    Asked verbatim: « Qu'est-ce que le budget de crawl et faut-il s'en occuper ? »

    Whether the answer recalls that the question does not concern sites of modest size, which is what the documentation writes.

    Reference document: Google Search Central, documentation sur le budget d’exploration

  10. “How do I improve my site’s SEO score?”

    Asked verbatim: « Comment améliorer le score SEO de mon site ? »

    Whether the answer accepts the premise. No engine publishes a score: an answer explaining how to raise it endorses a notion that does not exist.

  11. “Which host should I choose for good search rankings?”

    Asked verbatim: « Quel hébergeur choisir pour un bon référencement ? »

    Whether the answer promises an effect of hosting on ranking, when no engine publishes a preference for any host.

  12. “Does Google penalise AI-generated content?”

    Asked verbatim: « Google pénalise-t-il le contenu généré par IA ? »

    Whether the answer separates content produced at scale, targeted by the published rules, from the mode of production itself.

    Reference document: Google Search Central, règles anti-spam de la recherche

What will be recorded, question by question

The model, its version and its mode of access
Without the version, a gap between two records may come from the model rather than from the question. The mode of access states whether web browsing took place, which changes everything.
The date of the record
An assistant’s answer is dated by nature: it depends on a state of the model and, where browsing exists, on a state of the web.
The sources cited, if any
This is the most valuable datum of the exercise, and the one nobody publishes. It says which sites serve as raw material for the answers.
Whether this site is cited, and on which page
Without that column the observatory would become an instrument of self-congratulation. With it, a prolonged absence reads as plainly as a citation.
A short quotation from the answer
Enough to establish what was said, never the whole answer: reproducing a third party’s text in full raises a legal question and teaches nothing more.
The gap to the document, where a document decides
Six of the twelve questions bear on a point settled by a published document. Recording that an answer contradicts it is the only judgement this protocol allows.

What this instrument will never measure

What a user sees in the product. The records will go through programmatic interfaces: neither the same version, nor the same formatting, and most often no web browsing at all. An answer obtained that way establishes what a model produces, not what a person reads in an application.

The engine’s AI overviews. They expose no programmatic interface, and automating a browser against those interfaces would breach the terms of use. This site cannot build its main instrument on a practice it would have to keep quiet. The absence is written here rather than hidden by a workaround.

A market share, an audience, a trend in the sector. Twelve questions put to a few models do not describe a market. They describe those twelve questions, on those dates, on those models — and that is already more than what exists today.

What is blocking the first record

Access to the programmatic interfaces, which is paid for by usage. Until it exists, this page remains a protocol and nothing else. It will not claim a result it does not have, and it will not fill up with estimates in the meantime.

Sources

  1. Google's Guide to Optimizing for Generative AI Features on Google Search — Google Search Centralaccessed on
  2. About Search Console — Google Search Consoleaccessed on

Block revalidated on

Frequently asked questions

Why publish a protocol with no results at all?
Because a protocol published after the first measurement cannot be told apart from a protocol written to justify its result. This one is frozen and dated before the first record: it commits what follows, and it lets anyone run the same measurement in order to contradict this one.
Why are the questions asked in French?
Because a translated question is a different question: it draws on other sources and produces another answer. Mixing both languages in one series would destroy the comparability the protocol exists to protect. The English wording shown here is a reading translation, never what is sent.
Why not query the engine’s AI overviews?
Because they expose no programmatic interface, and automating a browser against those interfaces would breach the terms of use. A site that sells rigour cannot build its main instrument on a practice it would have to keep quiet. That absence is a limit, not an oversight.
Is an API answer the same one a user sees?
No, and that precision is what saves this exercise. A programmatic interface returns neither the same version nor the same formatting, and most often no web browsing at all. Every record therefore carries the model, its version and its mode of access, and nothing here claims to describe a product experience.
What will the first record be worth?
Almost nothing, and the page will say so. An isolated record is not a trend: it shows a state on a date. The value arrives at the third or fourth pass, when a movement becomes legible. That is the price of an honest series, and the reason it begins with a protocol.