Products
The SEO fixerSEO RecordsFixes the SEO and speed of your pages at the moment they are served.The SEO writerSEO AutopilotDescriptions, FAQs and internal linking written from your real data, and nothing else.The proofreader at the network edgeSEO OutpostThe same corrections, placed in front of your site: it stays free of any plugin.GEO measurementGEO SEO RecordsThe AI assistant channel, measured: reads certified by address, demand read, pages that pay.Measuring the impactEffect SEO RecordsYour site surveyed every month, plus the gap with the previous survey: what your fixes have achieved.Duo — SEO Records and SEO Autopilot: €339 instead of €398, for one year→See pricingLeader in SEO quality — with the figures to prove it
Four products measured, twenty-nine records, ten world firsts. A record, here, is a property measured under real-world conditions, dated, and replayable — via a published command or via the verification built into your admin panel. On that condition alone does the word hold.
World firsts
Each one is in the product, today, and you check it yourself. The day one of them disappeared, its line would disappear from here.
01
The corrector comes in two versions — server and browser — that render exactly the same page, byte for byte — backed by a public corpus of 77 executable cases verified across 20,000 pages. We are the only ones to publish a parity corpus.
sh outils/matrice-php.sh
02
Verification ships WITH the product: one click in your admin panel replays our records on your store — fabricated inventions caught and rejected live, your own figures, reversibility verified via SHA-256. We are the only vendor to ship its tests with its product.
bouton « Rejouer les preuves », dans le module
03
The verdict of your self-proof is sealed with Ed25519, and can be verified in your browser, at home, offline. A certificate on which you are as much the judge as we are.
seorecords.com/certificat
04
Every number and every standard code in generated text comes from your data: the outlier is named along with its unit, and generation stops and waits. Even the pairing matters — "100%" gets flagged when only "100 kg" is sourced.
compteurs de rejets, dans le module
05
Updates arrive via an Ed25519-signed, sequenced manifest: a replayed archive — even a correctly signed one — is rejected, and the SHA-256 fingerprint is verified before any write.
refus prouvé en production
06
After uninstallation, 3,993 files compared fingerprint by fingerprint: your site comes back to you 100% identical, database included. The module corrects the page as served, and your site keeps its state.
npm run preuve:installation
07
Every read by a declared crawler is checked against the IP ranges published by its operator at the moment it comes through, and the verdict is written line by line in the site's own log — with zero cookies and zero trackers — then joined to the pages that pay. The only AI assistant channel log that certifies what it counts.
npm run journal
08
When an AI assistant comes to read a page, a human has just asked a question on that topic, right now. Live reads, per page and per day, tell you what people come looking for on your site: real questions, read in your own log, at zero euros of API.
npm run demande
09
Equipping some pages, leaving others, and comparing certified reads before and after, with a permutation test on a published seed. The only assumption: the labels could have fallen differently. The result replays for anyone holding the log. The experiment requires hands on the site: we have them.
npm run essai
10
Every monthly statement is delivered to the merchant sealed with Ed25519: a dated record of their site's state, which they keep and which a third party can verify on their own machine, offline, in their browser: verification belongs entirely to them. Delivering evidence that stands up to scrutiny, which the customer can check themselves — we are the only ones doing this.
seorecords.com/certificat
Product 1 — the SEO corrector
The corrector takes over your pages at the moment they are served — and every number below comes from measurements taken on 13–14 August 2026, published with the command that produced it.
71% sooner
The same page served raw then corrected, in a real Chromium on a network throttled to 500 KB/s: complete in 806 ms instead of 2,791 ms — and LCP drops from 1,228 ms to 756 ms.
npm run preuve:vitesse
70× lighter
0.07 MB downloaded versus 4.9 MB (Yoast) — weights measured at the source. Installed: 228 KB and 30 files, versus 16,446 KB and 1,951 files.
node outils-js/comparatif.mjs
34seconds
Timed installation on a real WordPress, through the official channel — 7 fixes applied on the very first page served.
npm run preuve:installation
100% free at any time
3,993 files compared fingerprint by fingerprint (SHA-256): your site returns to exactly its previous state, database included. You stay free to leave whenever you want.
npm run preuve:installation
20 000tortured pages
Pages mangled at reproducible random, four invariants required on each one. 729,319 fixes applied — 100% cite their source. A thousand pages of torture replayed at every verification.
npm run fuzz -- --n=20000
586pages/s
Through the real middleware, end to end: median overhead of 1.12 ms per page. The corrector alone: 908 fixes applied to a 373 KB page in 15 ms.
npm run debit
100% on your side
Network capture in a real browser, plus a static test on all shipped code: everything stays on your side. Only one outbound call option exists, and you're the one who enables it — IndexNow, which notifies search engines that a page has changed.
npm run tout
2languages, same bytes
A public corpus of 77 executable cases, run through the server version (PHP 8.1 → 8.5, real binaries) and the browser version — which output exactly the same bytes, parity verified across 20,000 pages.
sh outils/matrice-php.sh
Product 2 — the SEO writer
The copywriting that publishes what it can prove, and only that. Its verification is built in: the "Replay the proofs" button measures these records on YOUR store, in a matter of seconds.
01
Every number and every standard code in a generated text comes from your data — the odd one out is named along with its unit, and generation waits. Even the pairing counts: "100%" is flagged when only "100 kg" is sourced.
02
Native writing is the default mode: deterministic (same facts, same text, down to the byte), 5 complete languages, everything happens on your side. In real-world conditions: hundreds of product sheets per second, average measured similarity of 8.2% across 19,900 pairs.
03
Every publication keeps its BEFORE state. Restoring returns the catalogue to exactly what it was — empty fields included, proven down to the SHA-256.
04
Every publication shows its clicks and impressions before/after when Search Console readings bracket it — labelled correlation, honestly.
05
Every internal link carries its reason: the default category, the declared brand, the active category neighbours. A link exists when the relationship exists in your catalogue — re-checked at publication, 100% live links.
06
You change a weight, remove a feature? Your published description has aged in perfectly good faith. The monitoring sees it, names it, offers the replacement — and steps aside when the real one comes back.
07
"2 years", "2 Jahre" and "2 ans" are the same canonical pair. A language edited on its own ("3 Jahre" on the German side) is caught and named, discrepancy by discrepancy.
08
Verification is built in: one click replays our records on your own shop — fabricated inventions caught and discarded live, byte-level determinism, your own figures, reversibility verified by SHA. Your catalogue comes back to you identical.
Product 3 — measurement of the AI assistant channel
The measurement counts who comes to read — visits, live reads certified by address, crawls, each in its place — and links the pages read to the euros collected. Every figure below comes from a dated reading, published with the command that replays it.
0 / 0false positive / miss
Pages from real sites frozen on 09/05/2026 with their provenance, in French, English and German: Shopify, WooCommerce and PrestaShop stores, press, government, encyclopedia, documentation. Ground truth declared on evidence, page by page, before being written: every code issued on a real page is earned, and the command replays it.
npm run conditions-reelles
5certified channels · 816 ranges
"ChatGPT-User" is a word in a header, and a word can be written. OpenAI, Perplexity, Apple, Microsoft and DuckDuckGo publish their crawlers' addresses: the address of every read is checked against those ranges as it passes, only the verdict is recorded, and the address is forgotten immediately. A certified read comes from its publisher's own network, and the log proves it line by line.
npm run plages
12 / 12in 403 · control 200
Reading taken on 09/05/2026 on nytimes.com: robots.txt blocks seven live-read crawlers, and the firewall returns 403 to the twelve AI assistant agents while the neutral agent gets through. The blocking targets them, and the measurement proves it: two layers, two figures, two causes — read in a single command.
node outils/analyser.mjs https://www.nytimes.com
0 €of API per month
Visibility platforms charge from 29 to 489 dollars a month because they pay the models to query them. We read the site's logs: everything is already there. One million log lines are re-read in 601 ms.
npm run records
44cases, two languages, 0 divergence
The PHP wrapper tables are generated from the Node version, and parity cases are computed by it then replayed in PHP. Tested on a WordPress installed in a subdirectory and on a PrestaShop 8. A table duplicated across two languages would be a truth held twice.
npm run php:tests
Product 4 — measuring the effect
This product reads your site every month and returns the differential against the previous reading. For a product that measures, only one question matters: did every reported change actually happen? These six figures answer it, and each one can be replayed.
100% of movements found
6,576 changes, checked one by one across 300 pairs of readings built by the very same computation that produces the reports served to customers. A "fixed" item must be present in the previous reading and absent from this one; an "appeared" item, the reverse; a signal that regresses must carry both exact counts; a page that has become clean again must have been flagged yesterday and no longer be. All 6,576 check out.
npm run records:effect
300 / 300pairs weighed in both directions
A computation that leaned toward good news would be a salesman's computation. On each of the 300 pairs, the delta is recomputed in the other direction: what is counted as fixed on one side must be counted as appeared on the other, and "regresses" must become "gains ground". The switch is exact on all 300.
npm run records:effect
900 / 900shuffles, a single verdict
Two readings of the same site list their signals in whatever order they come, and an order-sensitive computation would invent changes. The signals from both readings are shuffled at random 900 times, the differential is recomputed and checked against the original verdict — codes, counts, and named pages. The verdict holds all 900 times. This record failed at first: the reading of 07/09 showed 809 shifted verdicts out of 900, and sorting the pages before truncating them made it exact.
npm run records:effect
12readings delivered in 365 days
A full year of sweeps, simulated pass by pass — four passes a day — with the product's real decision-making: a thirty-day cadence, holding the reading in flight, the six-hour belt, and one reading in twenty deliberately left hanging. 12 launches, 12 readings delivered, one at a time. A reading that drags is retried, and the following month arrives on time.
npm run records:effect
1 460 / 12round passes · reads of the site
The sweep runs four times a day, 1,460 times a year. It requests a reading from the measured site twelve times: the other 1,448 passes look at the state of the reading on our side, and your site ignores them. A product that measures a site should cost it as little as possible — and this count says so.
npm run records:effect
100% of the figures in the email
300 emails are generated from 300 real differentials, then re-read: each of the 1,372 figures highlighted in their text is checked against the count it comes from. An email that inflated a figure by one would be caught here. All 1,372 trace back to their source.
npm run records:effect
100% of attested fields
The seal is the document you produce against a third party: what it certifies must be exactly what the difference contains. 300 certificates are sealed from 300 real differences, then re-read field by field — pages read, signals, fixed, appeared, declining, improving, pages back in good standing, and the difference page by page — changed, new, gone, moved. All 3,300 fields have their source. A certificate that embellished by a single unit would be worse than no certificate at all: it would be signed.
npm run records:effect
100% of the annual report fields
At the twelfth statement, the annual report goes out sealed: the twelve statements, and what changed from the first to the last — signals gone, arrived, declining, improving. 300 years of twelve statements are seed-generated, and each of the 7,200 fields in their reports is checked against an independent recalculation from the twelve summaries. All 7,200 have their source.
npm run records:effect
You've arrived from a Google ad. Do you allow us to tell Google if you subscribe, so we can measure our ads? Only your click ID and the amount are shared.