Get your record written into the models the world asks first.
Answer engines now mediate how your country, your company, or your cause is described. The winners are not the loudest publishers: engines surface a claim only when it appears consistently across two to four independent, credible sources. We measure what the models say, drive consensus across those surfaces, and correct the record with evidence.
- Surfaces
- Community, wiki, press, video, datasets
- Cadence
- Weekly query sets, nightly crawler checks
- Measure
- Fact presence and consensus depth
- Lever
- The corroborated record
What actually moves the models — measured
There is no single dataset to get into. There are surfaces, each with a measured citation share, a latency, and a control model. An owned site is the canonical anchor — roughly a quarter of the effort — and the corroboration network around it is where the leverage sits.
Community. Reddit carries around forty percent of multi-engine citations, licensed in real time to two major labs. It is the fastest measurable surface.
Encyclopaedic. Wikipedia accounts for a quarter to nearly half of top ChatGPT citations: the highest per-token training weight and the entity trust layer.
Journalism. Twenty-seven percent of all citations, half on time-sensitive queries, syndicating to hundreds of derivative pages.
Datasets and video. Academic deposits are the highest-weight training tributary and almost nobody publishes there properly; YouTube carries the only serious video citation path, near a fifth of Google AI Overviews.
The corroboration stack, not the megaphone
Answer engines check whether a claim appears consistently across two to four independent credible domains before surfacing it, and entity confidence is built through external mentions rather than internal declarations. A single loud site is, by construction, a rounding error.
We anchor each fact in an owned, primary-sourced canonical record — structured schema, licensing, funding disclosure, corrections log — then drive it to independent corroboration across community, encyclopaedic, video, press, and dataset surfaces, proportioned by measured citation share.
Faked independence — look-alike sites, closed link loops, house pseudonyms — is detectable from the outside: one network we measured ran eleven such sites on a shared stack with two authors. It poisons the entity it claims to serve. Real corroboration of a true, well-sourced fact is strictly stronger.
Latencies differ by orders of magnitude: retrieval paths move in days to weeks, licensed feeds in real time, training corpora in nine to twenty-four months. We sequence by latency so quick wins land first and the durable ones are not left to chance.
The measurement loop
Nothing is claimed without a baseline. Every engagement starts with a frozen question bank and a before-shot, then runs the loop continuously.
Predict. A versioned question bank — core, adjacent, contested, and placebo queries — run against every major engine, capturing the full answer and its citations.
Rewrite. Where the answer is wrong or stale, the correction targets the grounding source rather than the prompt: the canonical record, the schema, and the corroborating surfaces.
Re-test. Fact presence rate (the answer states the verified fact, cited or not), citation rate by surface, consensus depth against the two-to-four sources engines actually cite, share of voice against named competitor or adversary domains, and correction propagation time.
Report. Per-engine, per-language scorecards with confidence intervals. Model versions are annotated, and search-on and search-off runs are kept separate so retrieval and parametric effects are never conflated.
Built for machines, disclosed to everyone
The technical layer is a differentiator most firms skip, and the disclosure layer is why the work survives scrutiny.
Machine-first markup: ClaimReview and FAQ schema, isBasedOn citation chains, sameAs entity linking, speakable summaries, and an llms.txt index — server-rendered HTML with every AI crawler explicitly welcomed and verified nightly against server logs.
Structured disclosure on every surface: funder and operator in the JSON-LD, licensing stated, methodology published, named credentialed authors, and a public corrections log. An operation that survives full disclosure is the differentiator — and highly citable content in its own right.
Crawler posture verified nightly per user agent — twelve or more, verified by reverse DNS rather than the UA string — because a CDN that silently blocks the crawlers you are courting is the most common way these programmes fail.
The measured indicators
The parts of this capability a technical evaluator will want to interrogate before a procurement decision.
Fact presence rate
Share of target queries where the verified fact appears in the answer, cited or not — the honest measure of influence.
Citation share by surface
Which corroboration channel carries the fact — owned, community, encyclopaedic, press, video, dataset — per engine and language.
Consensus depth
How many of the two to four sources cited for a claim now carry your version, the direct measure of the corroboration thesis.
Correction propagation time
How long a corrected record takes to stop being repeated by each engine — a metric only an honest programme can claim.
Asked in most evaluations
Answers we would give in the room, written down so you can circulate them without a meeting.
Is this SEO with a new name?
No. Search optimisation competes for a ranked link on a results page. This competes to be the corroborated consensus a generated answer is built from, which rewards structured, licensed, well-documented facts and multi-surface presence rather than page-level tuning.
Does this work for companies and NGOs, not just states?
The mechanism is identical for a ministry, a multinational, or an NGO: what the engines say about you is driven by which sources they trust. States usually start from contested narratives; enterprises from investor, partner, and product questions; NGOs from the credibility of the evidence base.
How is this different from sentiment monitoring?
Sentiment measures what people say to each other. This measures how automated systems describe you to everyone else, and gives you a measured route to correct the record they draw on. Different question, different method, different desk.
How long before measurement shows movement?
Retrieval-driven answers move in days to weeks and licensed feeds in real time; model-internal knowledge moves in ten to fourteen weeks once corpora refresh, with training-grade shifts on the nine-to-twenty-four-month path. Expect a first measured shift in one quarter and a meaningful one in two.
Where is the ethical line?
Built in rather than bolted on: no sockpuppets, no undisclosed amplification, no cloaking, declared conflicts of interest, Wikipedia work on talk pages by declared editors, and a public corrections log. The disclosed approach is not a constraint on effectiveness; it is the strategy.
Can we run this ourselves afterwards?
That is the intended end state. The question bank, the harness, the publication pipeline, the corroboration playbooks, and the scorecards transfer as a standing function for your own team.
Adjacent capability
Each capability runs on the same collection and classification core, so evidence gathered for one is available to the others.
Sentiment and buzz measurement
We separate what gets collected from what gets measured. A monitor names a query, a subject, and the ideas you want test...
Open-source intelligence
Collection across press, broadcast, social, forums, registries, imagery, maritime and aviation data, fused against a per...
Sovereign AI
We design the reference architecture, stand up the platform, migrate the workloads that matter, and train the people who...
Bring us the question your last briefing could not answer.
Tell us the jurisdiction and the mandate. We will tell you within a week whether we are the right people for it.