Retrievability audit polyant.ai / docs.polyant.ai captured 2026-08-20 12 signals / 7 pages

Your documentation answers the question. It just cannot be quoted.

You measured that Polyant appears zero times across 115 non-branded readings. This is one hypothesis for why, tested against your own pages rather than against best practice. Every figure below is recomputed in your browser from HTML captured on 2026-08-20, and the assertion suite at the foot of the page re-derives all of it.

Prepared by Muhannad for the Exelab AEO/GEO pilot brief Independent, unaffiliated No client data used
01 / the finding

The marketing site is machine-legible. The documentation is not.

Your four stated buyer questions are about running on customer infrastructure, avoiding single-provider dependency, suitability for regulated European organisations, and governance with traceability and human oversight. Every one of those is answered in the documentation. The documentation is the least retrievable surface you own.

Scores are weighted sums over twelve signals, computed per page and pooled per host. A signal that does not apply to a page (comparison structure on a non-comparison page) is left out of that page's denominator rather than scored as a failure.

02 / reproduction

The zero reproduces, and the citations point somewhere else.

On 2026-08-20 I put one of your own buyer questions to a live answer engine, phrased as a prospect would: an AI agent platform, self-hosted, open source, enterprise, EU governance, an alternative to Copilot Studio. Nine platforms were named. Polyant was not among them.

What the answer cited

The sources behind that answer were third-party comparison and listing pages, not vendor sites: agent-works.ai, datanorth.ai, simplai.ai, ringly.io, kilo.ai. This matters more than the absence itself. For this class of question the retrieval surface is the comparison corpus, and you are not in it. Publishing a better page on polyant.ai does not, on its own, put you there.

The asset you already have

Your llms.txt is the strongest answer-engine asset on the estate. It states the positioning, names eight competitors with a differentiation line each, lists regulatory coverage, and it includes a "when NOT to use Polyant" section, which is the kind of self-limiting claim models quote readily. It is 2,669 bytes, it lives only on the marketing host, and its substance appears nowhere in the page HTML as schema.

03 / entity

The name collision has a name.

You reported that branded queries sometimes return an unrelated company. A live search for the bare token returns ten of ten results for Dimension-Polyant, a German sailcloth manufacturer founded in 1966, part of the Verseidag Group of Krefeld, with plants in Kempen and Putnam and a well-known X-Pac laminate line. It has six decades of entity history. polyant.ai does not appear on that first page at all.

This is an entity-resolution problem before it is a content problem. The competing entity is older, industrial, and unambiguous, so a model with weak evidence for a software company of the same name will reasonably prefer it. The counter is not more copy. It is a stronger, machine-readable identity claim: the product typed as software, bound to Exelab S.r.l. as publisher, with the sameAs edges that let a model separate the two.

Measured on your pages today

Three of your pages declare Organization. None of the seven declares SoftwareApplication or Product. So the vendor is typed and the software is not, on an estate whose entire visibility problem is that the software is confused with a fabric.

04 / evidence

Seven pages, twelve signals, every cell traceable to the capture.

Click a row to open its per-signal breakdown. Nothing here is stored as a number: the table is built by running the analyzer over the captured page records when this page loads.

PageHostScoreWeighted BandWeakest signals
05 / falsifiable

Run the same twelve signals on any page, right now.

A captured corpus is reproducible but unfalsifiable, so here is the same analyzer against a live URL. It resolves DNS first and rejects private, loopback, link-local and cloud-metadata addresses, revalidates every redirect hop, and caps the body it will read.

Ready. Try a competitor's comparison page for contrast.
06 / shape

What a six-week pilot tests, in the order it has to be tested.

The sequence is not arbitrary. Entity disambiguation has to settle before citation measurement means anything, because until a model can tell the software from the sailcloth, a change in citations cannot be attributed to anything you did. Full hypotheses, deliverables, dependencies and exclusions are in the proposal.

Weeks 1 to 2

Identity and baseline

Type the product, bind it to Exelab, and add the disambiguation edges. Re-read your existing prompt set unchanged so the comparison stays honest.

Weeks 2 to 4

Make the documentation quotable

Per-page descriptions, canonicals and question-shaped sections across the docs corpus, plus an llms.txt and a sitemap for that host. Highest measured gap, so it goes early.

Weeks 3 to 5

Enter the comparison surface

Turn the prose alternatives cards into extractable comparison structure, sourced from the differentiation lines your llms.txt already contains.

Weeks 5 to 6

Re-read and attribute

Second reading of the same prompt set. Report what moved, what did not, and which parts cannot be attributed to the work with the evidence available.

Six weeks is enough to change what is retrievable and to observe first movement. It is not enough to prove durable citation share. Any change measured this early is a signal to keep testing, not a result to bank, and I will report it that way. The honest limit, stated before the work starts
07 / verification

Every number above, re-derived and asserted.

A demo that prints a score is a claim. A demo that recomputes the score from its evidence and fails loudly on disagreement is a measurement. These assertions run in your browser now.