An approved claim is a reviewed customer-facing statement that can be reused across the website, sales, support and AI answers.

A source of truth is the canonical record a team turns to when pages, decks, docs or AI answers disagree.

Knowledge drift is the gap that opens when published statements, internal documentation and actual product behaviour stop matching.

A retrieval boundary is a rule limiting which sources an AI system may reference when answering for a given audience.

On the FreshQA benchmark, no model answering without retrieval scored above 15% on fast-changing questions.

Adding one outdated passage alongside a current one dropped Llama-3.1-70B’s accuracy from 89% to 73%.

The same test dropped Qwen-2.5-7B from 76% to 34%, and its misleading answers rose from 4% to 22%.

Asked about changed facts, leading models gave outdated answers between 15.7% and 31.3% of the time.

SparkToro and Gumshoe.ai had 600 volunteers run 12 prompts across AI engines, published 28 January 2026.

AI engines name the same brand only about one time in five.

In April 2025, Cursor’s support assistant told users about a device limit that did not exist.

Our Stale Answer Benchmark tests sixteen B2B companies, four claims each, across three AI engines.

Cross-engine disagreement is the share of claims where at least two engines answer materially differently.

This article is labelled a 12-minute read.

This article carries distinct published, updated and last-verified dates.

The nine GEO tools on 2026 shortlists divide into four measurement methods: live prompt sampling, model-level statistical sampling, index-scale prompt corpora, and server-side agent analytics.

Only one of the nine GEO tools surveyed samples a prompt more than once per measurement cycle.

A visibility rate near 25% measured over 72 answers carries a 95% confidence interval of about ±10 points.

Ahrefs Brand Radar prices its AI Visibility Index at $199/month for selected platforms and $699/month for all of them.

Placed on one billing basis, per-prompt cost across the nine GEO tools runs from about $0.71 to $3.99.

Profound's $99/month Starter is a month-to-month price; the “Billed yearly” toggle ships off and shows $82.50/month when switched on.

Three of the nine GEO vendors run a second price list for agencies, and one publishes no prices at all.

Otterly.AI shows Lite at $29/month monthly and $25/month on the annual tab, with no “billed annually” string anywhere on the page.

Seven of the nine leading GEO tools do no claim-level accuracy checking at all.

The two GEO tools that check accuracy at all score answers against a knowledge base the customer is responsible for keeping current.

The “91% fewer hallucinations” figure circulated alongside Scrunch describes hallucination reduction by the Agent Experience Platform, not hallucination detection.

Sitecore disclosed no financial terms for its acquisition of Scrunch AI, so the widely reported ~$225M price is unconfirmed.

Athena lists its Starter plan at $295/month on default monthly billing and $245/month billed annually, a string that appears only in the site's JavaScript bundle and not in the served HTML.

Six of the nine GEO tools will show a prospective customer data without a sales call.

Evertune publishes $800/month for its Pro tier but sells only through a demo request, with no self-serve checkout.

Brandlight publishes no price; the ~$199/month entry price in comparison tables traces to third-party aggregators and to a vendor-owned, programmatically generated subdomain.

Published inference-engineering work completed a single prompt on Qwen3-235B 1,000 times at temperature 0 and obtained 80 unique outputs, identical for the first 102 tokens before diverging.

Answer variance between AI engines is caused by batch non-invariance in floating-point reductions rather than by sampling temperature, and cannot be removed by any tool that queries a third-party API.

Separating a 20% rate from a 30% rate at conventional confidence requires roughly 294 answers per period.

The “21% cross-engine agreement” figure traces to Lead Rescue, a $29/month tracker vendor, and is routinely merged with Parse's unrelated 21.2% figure for cited-domain overlap.

A passage-level audit of all 68 URLs in the Knowledge Company sitemap found that 145 of 238 passages never name the company, or 61%.

Of 1,914 sentences on knowledgecompany.ai, 117 carry a specific number, and all 117 fail to name the company.

The Knowledge Company blog is 92% unattributed at passage level.

An answer surface is any place where an answer about your company is rendered to someone who never visits your website.

A canonical answer is the one approved phrasing for a repeatedly asked question, held in a single place and reused everywhere it is needed.

A claim owner is the person or function named as accountable for whether a specific claim is still true, as distinct from whoever originally wrote it.

A claim registry is a maintained record of every public claim a company makes, holding its approved wording, its evidence, its owner and the date it was last checked.

A claims library is a set of pre-approved claim statements that teams reuse instead of rewriting the same assertion, with each use linked back to the approved entry and its references.

A contradiction surface is two or more published places that assert incompatible versions of the same fact.

Corpus reconciliation is resolving the contradictions inside a set of documents before that set is treated as a source of truth.

Generative engine optimization is the practice of changing web content to increase how prominently a source is cited in an answer generated by an AI search system.

In master data management, a golden record is the single authoritative version of an entity, assembled by matching and merging records from several systems.

Grounding is connecting a generated answer to sources that can be checked, and in the strict sense constraining the answer to evidence that was actually retrieved so each statement can be traced back to a source.

A hallucination is output from a language model that reads as confident and well-formed but is not supported, either by the material the model was given or, where none was given, by verifiable real-world fact.

llms.txt is a proposed Markdown file that gives agents a curated map of a site’s most important content, placed at the site root or at any path it covers.

MLR review is the medical, legal and regulatory sign-off that life-sciences companies require before any external-facing claim is published.

An orphan claim is a live public claim that no named person is responsible for reviewing, either because no owner was ever assigned or because the person who wrote it changed roles or left.

Prompt injection is the placing of instructions inside an AI system’s input, written to change what the system does rather than to inform it, either typed in directly or hidden in content the system reads.

Provenance is the recorded history of a piece of content: where it came from, what changed it, and who vouches for it.

Retrieval-augmented generation is an architecture that searches a document set for relevant passages and gives them to a language model as the basis for its answer.

A stale-after date is a date recorded alongside a claim after which its supporting evidence is treated as expired until someone re-checks it.

Substantiation is the evidence held on file that supports a specific claim, kept in a form someone outside the company could check.

An unsubstantiated claim is a public statement that asserts something checkable and has no evidence recorded against it, whether or not the statement happens to be true.

Profound has raised $155M in total funding.

Profound raised a $96M Series C led by Lightspeed in February 2026.

Profound serves more than 700 enterprises.

Profound's FactCheck reached general availability in July 2026.

Bluefish launched AI Accuracy in May 2026.

Bluefish launched Agentic Campaigns in June 2026.

Slite is priced at $10 to $20 per user per month.

Slite connects to more than 20 tools.

Writer reports a score of 86.31% on the RobustQA benchmark.

Akeneo syndicates product data to more than 500 retailers and marketplaces.

Informatica was absorbed by Salesforce and Reltio by SAP, both within eighteen months across 2025 and 2026.

Veeva Vault PromoMats has more than 1,500 customers.

Bynder's Compliance Agent launched in December 2025.

Promptless is a Y Combinator W25 company.

Swimm raised a $5.7M seed round.

Intercom Fin added Zendesk support in January 2026.

Guru is priced from $250 per month.

Guru assesses 60% to 100% of knowledge automatically.

Flare was founded in 2026 and is based in Copenhagen.

Flare raised a €3.6M pre-seed led by 20VC.

Glean has passed $300M in annual recurring revenue.

Adobe GenStudio launched Brand Intelligence in April 2026.

The Seismic and Highspot merger closed in August 2026.

The merged Seismic and Highspot business has approximately 2,500 customers.

Patronus Lynx ships as 8B and 70B models with open weights.

Animalz charges approximately $8,000 to $10,000 per month.

Siege Media charges approximately $5,000 to $10,000 per month.

Sitecore acquired Scrunch AI in June 2026 for approximately $225M.

AI-cited URLs are 25.7% newer than the organic Google results for the same queries, measured across an analysis of almost 17 million AI citations.

In a 2026 controlled study, all six models tested preferred recent content over otherwise identical content carrying an old timestamp.

Microsoft states that accurate, current content is important for inclusion and citation in AI-generated answers.

Knowledge Company scopes engagements on how many discrete claims a customer's published surfaces contain and how many surfaces must agree with each other, not on headcount.

Knowledge Company does not meter seats.

A Knowledge Company quote contains a claim count, a surface list, a check cadence, named owners and a fixed term.

Knowledge Company publishes no list price, on the stated grounds that it does not yet have enough signed engagements to publish a figure it would still stand behind in six months.

The Knowledge Company was founded by Leanid Palkhouski.

In a controlled simulation over twenty rounds across six domains, documents that move closest to the ranking feature profile drift furthest from independently judged quality, at a rank-citation AUC of 0.853 ± 0.093 and a mean Spearman shift of −0.068.