Changelog
One entry per milestone shipped. Nothing is announced here before it runs in production, a changelog that promises is the fastest way to collect complaints.
Latest entry:
-
One broken link is one finding, however many pages carry it
An outbound broken link was recorded at the grain of the source page: one failing target produced as many rows as there were pages linking to it. For a link sitting in a footer, that meant thousands of twin findings, a screen repeating the same sentence, and above all a health score collapsing, since every page on the site ended up marked "in error" over a link the site does not control. On an eighteen thousand page site the analysis chain broke outright: accessibility, GEO readiness, internal linking and the summary stayed empty for pages that had already been explored.
The finding now lives at the grain of the target, with the number of pages referencing it. Across audits already on record, one site regains more than forty points of score and the others do not move: a fix that only gives back what it had taken away.
In the same pass, "a page in error" stops existing as three separate definitions that nonetheless published the same score, and a test now compares the three surfaces against each other.
Crawl
-
AI Overview and AI Mode citations, and access read from the server
For every tracked keyword, the daily reading now reports citations in AI Overviews and AI Mode: the domains cited, the ones showing up in your place, and the share of your tracked queries where your site is cited. The two surfaces are never added together. One query can cite five sources in an AI Overview and none in AI Mode: two denominators, published side by side, and neither is a fallback for the other. The rank tracking workbook carries the reading keyword by keyword, and the PDF report publishes the summary.
The module measures your tracked queries, not your market, and that sentence sits on the screen rather than in a footnote. A day without an AI Overview is not a "not cited": it is a measurement that did not happen, counted as such. A signal warns of a real loss of citations, and stays quiet when the overviews themselves have retreated, so your site is not blamed for the engine's own move.
The AI visibility table, for its part, read the access declared in robots.txt and the presence of crawlers in your logs, never the status actually served. A "Served" column now publishes refusals with their denominator, including when the verdict stays healthy, and two states are added: a crawler allowed on paper but turned away by the server, and a crawler the active probe cannot reach where the logs say nothing.
The access axis of the GEO score no longer rests on robots.txt alone: the real reachability check runs on every campaign. The scale is unchanged; a branch that was already written simply opens.
Visibilité IA
-
A suggested site profile, with the sentence it came from
When a project is created, the home page and a second page picked without randomness are digested, then handed to a model that suggests the sector, the business model, the language and the country. Every suggestion carries the quote from your site it rests on, and "read on the page" stays distinct from "suggested".
Nothing inferred lands in your brief on its own: the screen suggests, you click, and it is your click that writes. A machine's guess must not become indistinguishable from a declaration, on the one piece of data whose worth comes from where it came from.
Three outcomes, never two: a profile filled in, a site with too little signal to conclude, or a collection still under way. A failure on our side is never displayed as a lack of substance on yours.
Prise en main
-
Every URL now says what each source knows about it
An audited page now carries three crossing panels: what the crawl found, what the server logs saw go by, what Search Console reports. They load after the screen, so they never delay it, and a module that is not connected says exactly that instead of showing a zero in its place.
The distinction holds all the way down: "not connected" offers the missing step, "nothing for this address" is a fact and offers none. Rankings and AI citations appear as links rather than numbers, because they describe a query and a domain, not an address.
The path runs both ways: the tables in the crossing reports lead to the dossier of the address they diagnose, and search accepts a pasted address, including one never crawled, one carrying a different host than the stored one, or one reduced to a bare path, the only form logs ever carry.
Audit de site
-
Precomputed projections, and a guard that refuses a slowdown
Screens that aggregate a lot of data no longer rebuild everything on every view. Internal linking and server logs are projected at crawl time, then read at exactly the grain needed: the internal linking screen goes from over three seconds to roughly two hundred milliseconds, and the frame of two crossing screens paints in about fifteen milliseconds instead of several hundred.
Whatever still needs computing arrives after the screen, under a skeleton shaped like the content you are waiting for, never behind a blank page.
A guard measures these screens on every integration run, against written budgets: a slowdown is refused before it reaches production instead of being noticed there.
Performance
-
Placement, anchors, and a crossing with the audit
The daily snapshot carried eight fields no screen ever read. They are published now: where the link sits in the referring page (footer, body, article), the declared nature of the links, how concentrated the network of addresses and subnets behind the referring domains is, the age of the profile, and inbound links pointing at a page that is gone.
Placement changes the story. A profile summed up as "100% Dofollow" can be, in practice, overwhelmingly made of footer credits. A badge that holds true of every case informs no one; the breakdown does.
Anchor text is characterised (brand, bare address, no anchor at all, other), and an empty anchor is stated as a fact rather than left as the first row of a table sorted by volume. The link profile is also crossed with the pages the audit actually explored.
Backlinks
-
NessFlow opens an invite-only beta
Access is by invitation: open sign-up is closed, and a team already with us can still bring one of its members in. This phase exists to put the product in front of sites nobody here chose, not to show it to more people.
We are after the sites that put an audit tool in the wrong: JavaScript-rendered content, multiple subdomains, multilingual setups, large volumes, edge protection. What we are looking for, and what this phase does not promise.
Bêta
-
Creating a project, preflight, and getting started guide
A project can be created at sign-up, from the project switcher or from a dedicated screen: one definition behind all three paths, so one outcome. Creation leads to project setup, not to a page with nothing on it yet.
Before the first crawl, a preflight checks that the site answers us and names what stands in the way. Server log upload runs through a wizard, and a getting started guide follows the steps actually reachable on your plan.
Prise en main
-
Stopping a crawl, clearing a stuck one, and crawl rate
A running crawl now stops from the screen, right away: the order is carried out on demand instead of being queued, where a paused queue could leave it sleeping next to the work it was meant to interrupt.
A crawl whose execution has vanished releases itself after a delay, so the project no longer stays blocked. And the politeness delay between two requests is now genuinely applied: the crawl deadline accounts for it, so a fragile site is crawled slowly instead of being dropped halfway.
Crawl
-
Security signals, passively detected
Passive, non-intrusive detection of the signals that weigh on your visibility: exposed sensitive files, end-of-life technologies, TLS and headers, unencrypted resources on a secure page, Safe Browsing status. Every finding is stated in plain business terms, with its impact on search.
Nothing starts without an authorization signed by an administrator, versioned and kept on record: it alone permits a check, and a new version calls for a new signature.
Signaux de sécurité
-
Internal linking, at crawl and project level
Two new screens: the structure of a crawl (depth, page clusters, orphan pages) and a project’s internal linking tracked over time. Excel tabs and a PDF section follow.
Link Scores are ranks, not grades: they order your pages against each other and do not enter the health score. A page with no internal inbound link is a fact; a page whose computation did not complete is not filed with them.
Maillage interne
-
Keyword research and AI intent
A single reference where the tracking facet, which spends, sits next to the semantic facet, which does not: clustering by meaning, winnability estimates, search volumes.
New "AI intent" view: share of citations captured across three providers, with its denominator shown next to it.
Mots-clés Visibilité IA
-
Daily rank tracking and backlink profiles
Daily rank capture at full depth, with SERP features, and competitors extracted from captures already paid for rather than from a second call.
The backlinks module requires explicit activation by an administrator: an inactive project spends nothing. Profile, lost and gained domains, on-demand gap analysis.
Positions Backlinks
-
Project-level recommendations, and scorecard
Recommendations are no longer issued one per finding: they cross a project’s modules (crawl, server logs, Search Console, AI visibility) and come out ranked by impact and effort, each with the signals it rests on.
The dashboard carries a per-pillar scorecard, where a pillar with no data stays empty rather than counting as a zero.
Recommandations IA
-
Rendered accessibility audit, and conformance report
The accessibility audit runs against the rendered page, not the initial HTML; the distinction that decides everything on a modern application.
Contrast cards with a suggested fix, screenshots of failing elements, and an exportable conformance report in the format public-sector and US buyers ask for.
Accessibilité
-
Sovereign log analysis and Search Console cross-referencing
Server log upload up to two gigabytes, a streaming parser covering thirteen formats, and a human-versus-crawler traffic split.
No IP address is stored or sent to a model. Uploaded files are deleted after processing.
Three sources now meet on the same URL (crawl, logs, Search Console) with a diagnosis when they disagree, and orphan page detection.
Logs Search Console
-
Page-level semantic optimization for a keyword
For a keyword and the page targeting it, the analysis compares the expected search intent with the one the page actually serves, measures the semantic field it covers and suggests target occurrences.
It also recommends the title, H1, meta description and subheadings, and flags the pages in your project competing for the same keyword.
Optimisation sémantique
-
Evolution tab, and regression sentinel
A project keeps the history of its comparable crawls. Per-crawl series and daily series do not share the same grain: they live on two separate charts and are never added together.
The sentinel compares each new crawl against the project history and warns once per anomaly; everything else goes into a daily digest.
Évolution
-
Full Excel export, and executive PDF
Every results screen exports all of its data to a multi-sheet Excel workbook: issues, pages, links, recommendations. The workbook is streamed row by row, so its size does not depend on available memory.
A crawl page also produces an executive PDF with conditional sections: a section only appears when the data behind it actually exists.
Exports
-
Per-crawl "GEO readiness" report
Every crawl now produces an answer-engine readiness report: an AI crawler access matrix read against RFC 9309,
llms.txtvalidation, and opt-in active reachability checks.The score is versioned: changing the scale does not rewrite history, it opens a new comparable series.
Visibilité IA
What is missing from this page, and why
A module whose code is written but whose rollout is not done has no entry here. That covers billing: its build is closed, its operation is not. Listing it would date a promise rather than a delivery.
Start with a measurement, not a promise
Run an audit on your own site and read what the engine finds. If you would rather be walked through it, book a demo: we run it on your site, with your own URLs on screen.
Access opens in waves: we email you when yours is ready.