The demonstration, not the promise
One SEO prioritisation question, asked twice. First to an assistant with no access to the data. Then to Iris, connected to the technical audit and the Search Console of the website. Both answers are reproduced as they came.
Short answer
A generic AI assistant answers from general knowledge. A connected assistant first reads the data of the website: technical audit, Search Console, pages actually online, competitors. The difference is not the quality of the advice, but its order. Connected advice names a page, a query and an action. Generic advice names a best practice.
The test
The website used for this demonstration is Atelier SEO itself. Its technical audit, its content plan and its Search Console figures are public, so every claim below can be verified without taking this page at its word.
The question asked
On this website, which page should be worked on first for search?
Start with your pillar pages. Check title tags and meta descriptions, improve loading speed, strengthen internal linking. Publish content regularly and monitor Search Console.
The advice is sound. It would be identical for any website, which is exactly its limit: it does not say where to start on this one.
Three pages cover the same "SEO audit" intent: the home page, /audit-seo and /audit-seo-site. They will compete for the same queries as soon as impressions begin. The home page keeps the role of general entry point. /audit-seo should document what the ninety checks actually contain, the only defensible difference from the home page. Search Console still returns no clicks over ninety days, so the success criterion has to be measured in absolute terms at the first appearance of data, not as growth.
This answer names three precise addresses, a conflict between them, a role for each page and a measurement criterion suited to a website with no history. None of it can be deduced without reading the site and its data.
What follows belongs to a human, and it diverged. Rather than keeping three separate pages, the team consolidated /audit-seo-site into /audit-seo on 7 August 2026. The assistant detected the conflict and framed the trade-off; the decision to merge was taken by a consultant. That consultant has a name and a track record: who writes the analyses published here.
Both sources quoted here can be consulted: the content plan of the website, the public results journal.
The comparison
Both answers address the same question. The table does not compare their style, it compares what each one is able to name.
| What the answer contains | Assistant without data | Iris connected to the site |
|---|---|---|
| Pages named | None. "Your pillar pages" describes a category. | Three exact addresses, two of them in conflict. |
| Source of the diagnosis | General search engine optimisation practice. | Crawl of the live pages and Search Console readings. |
| Order of work | A list of tasks, with no priority between them. | One role per page, so an order can be deduced. |
| Success criterion | Monitor Search Console. | Measured in absolute terms, because the site starts from zero clicks. |
| Verifiable by the reader | Nothing to verify, nothing is specific to this site. | Every finding points to a page or to a dated reading. |
The generic answer is not wrong. It is true for everyone, which makes it unusable for deciding what to do on Monday morning.
The mechanism
Three recommendations from the example above are out of reach for an assistant with no access to proprietary data. They do not require better reasoning, they require a reading.
Cannibalisation cannot be deduced from a topic. It is observed by crossing the pages actually published with the queries they appear on. Without the crawl and without Search Console, the conflict stays invisible.
A website with zero clicks over ninety days cannot be judged on a growth percentage. Iris reads that zero and switches the measurement to absolute values. An assistant without data proposes a percentage increase, which means nothing from zero.
Recommended internal links are checked one by one against the list of addresses actually online. That is what stops an action plan from sending a writer to link to a page that was never published.
The objection
The question deserves a direct answer, because a generic assistant costs less and replies faster.
The limit is not the power of the model. A generic assistant has no access to the proprietary data of a website: not its crawl, not its Search Console, not its ranking history. It therefore cannot know which of two pages already receives impressions, nor which of the two deserves to be kept.
As long as the question is about method, a generic assistant is enough. As soon as it is about this specific website and the order of the work, it can only answer at the level of the category. The difference is paid in time lost on the wrong tasks.
Setting it up
Iris is a layer connected to an existing website, not a tool installed inside it. Nothing is added to the website code, and no migration is required.
The website is crawled the way a search engine would, from its public address. No server or CMS access is needed to produce the first diagnostic.
Search Console and Analytics attach through a Google account, in read-only mode. The connection can be removed whenever you decide, and the permissions stay visible on the Google side.
The action plan is read and applied in the CMS already in place. It assumes no change of technology and no dependency on the tool in order to publish.
Once connected
The description below is functional. It covers what can be observed in the workspace, with no promise of results.
More than ninety checks across four families, replayed on every new audit, with the affected URLs attached to each finding.
Positions, impressions and clicks by page and by query, on a rolling window, kept separate from estimates.
Findings are ranked by their observed effect on indexing, clicks or speed, rather than by a generic scale.
Every task comes with its target page, its search intent, its verified internal links and its measurement criterion.
To go further: what the audit checks in detail, how the agency works, the support formulas.
Your baseline
Run the audit, connect Search Console, then ask the question that concerns you.