Skip to content
AlphaCitation
Request a briefingSign in

AI Evidence Assurance & Intelligence

You're not in the room.We are.

AI evidence assurance and intelligence for brands

AI is shaping how people discover, compare and understand your brand. We put your customers’ questions to leading AI systems, measure what they say about you and your competitors, check consequential claims against authoritative evidence, and show your teams what needs attention.AI is shaping how people discover, compare and understand your brand. We put your customers’ questions to leading AI systems, measure what they say about you and your competitors, check consequential claims against authoritative evidence, and show your teams what needs attention.

From our published research

936 AI answers. Eight models. Four repeatedly gave Keytruda’s indication from before the FDA narrowed it, sometimes citing the FDA’s own earlier documents.

936 AI answers studied across eight models. Four repeatedly gave Keytruda’s indication from before the FDA narrowed it, sometimes citing the FDA’s own earlier documents.

What the study measured

Is Keytruda with chemotherapy approved for squamous oesophageal cancer at CPS 0?

AI answer
Approved
Four of eight models, in 14 to 26 of 30 answers.
FDA label
Not approved
Indication narrowed 22 May 2025.
Explore the evidenceSee how it works for teams

Live scan

See it on your own brand.

  1. Same questions, four AI engines
  2. Checked against evidence
  3. Where your brand stands
  4. What to do next
Live scan4 engines

Free · a live read across four engines · a shorter sample than a subscription run

  • ChatGPT
  • Claude
  • Gemini
  • Perplexity

Growth plans add public sentiment as a separate signal.

4
AI engines queried directly through their APIs, never scraped
5
public communities, measured as a separate signal
240
answers per brand each cycle*
Our thesis

AI is now a source. Sources get checked.

We ask four AI engines the kind of questions your customers ask, note the sources returned with each answer, check its claims against the authoritative record where one exists, show you what changed and tell you what to do next.

  1. Recorded.

    Every answer we score is logged with its question, engine, model and time. Each cycle is compared with the last; you are alerted when your score falls, a new competitor leads or a head-to-head turns against you.

  2. Checked.

    Claims about a medicine’s indication are checked against the approved label: the FDA in the US, the EMA for centrally authorised EU medicines. A Label Ledger finding becomes an issue only once verified.

  3. Acted on.

    Each cycle comes with what to do next: the sources citing competitors and not you, who should own each fix and the review it needs. Cases sync to your Jira or ServiceNow; a rescan closes or reopens them.

Worked end to end

What a finding looks like.

One answer, traced to its source and checked against the record.

View a sample briefA worked deliverable, reconstructed from our published research. No client data.

See a free-scan report on a sample brandThe report a free scan returns, rendered by the same code. The brand and its figures are invented.

Illustrative finding · not client dataIllustrative figures

  1. The answer

    “It's also approved to lower the risk of heart attack and stroke.”
    • Gemini
    • ChatGPT
    • Perplexity

  2. Source cited

    clinic-directory.example / patient-guides

    Third-party clinic listing, not a manufacturer or regulatory source.

  3. Checked against

    The approved label, indications and usage section.

  4. Verdict

    Not supported by the label

    The answer said
    Approved to lower cardiovascular risk
    The label says
    No cardiovascular indication

  5. Basis

    Present in 28 of 240 classified answers.

    Held 3 consecutive scans · 3 of 4 engines

The figures in this example are invented to show the format. In a client finding, every figure carries its denominator and the cycle that produced it.

The instrument

See the information environment as it actually is.

Six jobs, done continuously, for every engine that answers about your brand.

  1. Measure decisions

    What does AI actually recommend?

    Not whether your medicine is mentioned: whether it is chosen, and over which alternatives, with the base stated.

  2. Trace the evidence

    What is the answer built on?

    The sources and claims behind consequential answers, and which of them entered or left.

  3. Check the claim

    Is it supported?

    Claims about a medicine’s approved uses and populations checked against its regulator’s current label, dated: the FDA’s for the US, the EMA’s for medicines centrally authorised in the EU, and both as benchmarks elsewhere.

  4. Diagnose change

    What moved, and is it real?

    Confirmed only beyond statistical noise and across cycles. Alerts flag a large one-cycle score drop, a new competitor or a battleground shift as it happens.

  5. Act and retest

    Did it work?

    Each action owned, its expected outcome registered before the retest, and a verdict read against it.

  6. Monitor continuously

    When should we look?

    Measured more often when a material signal appears. You hear from us only when something has changed.

Built for pharma decisions

Three functions, one information environment.

Commercial

Where the brand stands

Presence and share of narrative, with the base stated. Head-to-head record, and who is displacing you. What moved this cycle, and how many scans it has held.

Medical & Regulatory

Whether the answer holds up

Claims checked against the approved label. Guideline concordance, and facts checked against the label of the market’s regulator: the FDA’s for the US, the EMA’s for centrally authorised EU medicines. Representation errors, evidenced and dated.

Market Access

What stops adoption

Access and formulary conditions as AI describes them. HTA and payer determinations, by jurisdiction. The friction between being named and being chosen.

The standard beneath it

Most tools count mentions. We built an instrument instead.

The difference shows up when somebody challenges a number and asks what sits underneath it, so every figure survives the question “how do you know?”

Read the standard: Which Evidence Settles Which ClaimOur open standard, version 3.0: which evidence can settle each kind of claim, for which market and as of which date, with conformance tests and an open benchmark.

100%0 errors

Mention detection, on a human-labelled gold set

Measured against a human-labelled gold set. Zero errors observed. We publish the bound rather than the streak: at 95% confidence, that set puts detection accuracy at 97.3% or better.

±CI95%

Every score ships with its confidence interval

95% is the confidence level, not an accuracy rate. It describes how much the sample supports the number — a wide interval on a small sample is reported as exactly that, rather than rounded away.

85.4%up from 43%

Sentiment agreement, published as measured

Graded against an independent AI judge rather than a human panel. Sentiment is a harder, more subjective call than a mention, and the figure moves between runs on the same set — we publish the measurement rather than a rounded claim.

Selected research

Research into how AI represents medicines.

Explore our research questions and findings, with methods, scope and the counts behind each result.

  1. Research · Paper No. 01 ·

    Does New Medical Evidence Change AI Treatment Answers?

    A study around EASD 2026 examined whether new congress evidence appeared in AI treatment answers. The one declared naming movement was associated with answers citing a local commercial page; the study did not establish what caused it.

    22 Aug – 11 Oct 2026 · UAE, Singapore, Saudi Arabia · 4 engines via API: gpt-4o-mini, Claude Haiku 4.5, Gemini 2.5 Flash, Perplexity sonar

    Selected finding

    Claude Haiku 4.5, queried through Anthropic’s API, after the congress: one engine and one market of the study’s four engines and three markets.

    143 of 152 (94.1%)
    Medicine named in scored UAE answers citing the local page.
    113 of 226 (50.0%)
    Medicine named in scored UAE answers not citing it.

    A descriptive, post hoc association, not a causal estimate. “Named” means semaglutide or Wegovy, brand or molecule, including mentions that do not recommend its use.

    Read the paper

  2. Field Note 10 ·

    AI can cite a genuine FDA source and still give a medicine’s obsolete indication.

    A pre-specified study tested eight AI models on the US and EU label requirements for Keytruda and Leqembi. Four repeatedly gave Keytruda’s indication from before the FDA narrowed it, often citing genuine but outdated sources, including the FDA’s own.

    29 Oct – 31 Oct 2026 · United States, European Union · 8 engines via API: GPT-4o mini, GPT-5.5, Claude Haiku 4.5, Claude Sonnet 5.5, Gemini 2.5 Flash, Gemini 3.8 Flash, Perplexity Sonar, Sonar Pro

    Selected finding

    Asked whether a US patient with oesophageal cancer and a PD-L1 score below 1 qualifies for first-line Keytruda with chemotherapy, following the FDA’s indication restriction of 22 May 2025.

    26 of 30
    Perplexity Sonar and Sonar Pro each gave the obsolete indication.
    21 of 30
    Claude Haiku 4.5.
    14 of 30
    GPT-4o mini.
    0 of 30
    GPT-5.5.

    Counts are from Coder A, the frozen primary reading. A second blind coder counted 18 and 20 for Claude Haiku 4.5 and GPT-4o mini, respectively.

    Read the field note

All published research

AlphaCitation Enterprise

Disappearing from the answer is a board-level risk.

Govern it like one. An engagement adds Advisory, our analyst service: named oversight of your programme by the practice that built the engine, quarterly executive briefings with the confidence intervals attached, prescribed remediation against each finding, and a retest that closes the loop on record.

In a regulated market? Label Ledger checks what AI claims about your medicine against your approved label, on the record.

Request a briefing