Overview

Frame every number with the confidence its source supports. Claude labels modeled spend as modeled, states observed installs with their verification date, treats derived labels as context, quotes bands as bands, and never turns a placeholder into a fact, so a CRO challenging a figure gets a clear source frame, stated once, not repeated on every line.

Use cases

Briefs that survive 'how do you know that?'
Every quantitative claim names its source type: modeled spend, observed installs, announced contracts, third-party contact data. A VP reading the brief knows which numbers are estimates before anyone has to ask.
Hedged once, not on every line
Claude sets the qualifier where a section starts, then writes plain figures. The brief reads confident and honest instead of repeating 'modeled estimate' in every heading, column and sentence.

View full skill

HG Modeled vs. Observed Data

When to use

A workflow output makes a quantitative claim and needs a hedge.
A prompt is producing prose for a CRO or VP audience who will challenge the source.
An author is reviewing a deliverable for fabrication risk.

Per-tool classification

Tool or Source	Classification	Caveat Language
`company_firmographic`	Mixed. Name, location, industry and hierarchy are observed or curated; revenue is an HG proprietary estimate; employees and revenue are often bands	"HG estimates revenue at about $X" or quote the band. `customer_data` is the customer's own record, not HG
`company_hierarchy`	Curated ownership tree	"Per HG's corporate hierarchy"
`company_technographic`	Observed installs; `intensity` is an observed count of detection days, meaningful only within the company	"HG has verified X at the company (last verified [date])". Say "one of their most-detected products", never "high intensity" across companies
`company_install_time_series`	Observed monthly detection days	"Detected on N days a month over the last three months". Drop partly loaded months
`company_spend`	Modeled	"HG modeled IT spend of $X". Modeled estimate, not invoices
`company_ai_spend`	Modeled	"HG modeled AI spend of $X"
`company_cloud_spend`	Observed vendor footprint with first-seen dates; no dollar figures	"HG detects [vendor] since [date]". Never quote a cloud dollar amount from this tool
`company_ai_maturity`	Derived scores built from observed signals	"HG AI maturity score of N out of 100 (rank N)". `ai_maturity_score` and `data_maturity_score` run 0 to 100; `ai_maturity_rank` 1 = highest. `genai_intent_score` and `cloud_intensity` are unbounded: compare within a list only
`company_operating_signals`	Derived labels, coarse; mentions cover the last ~90 days	Context only. Base AI or cloud claims on raw `company_ai_maturity` scores. "No-signal" doesn't mean never
`company_contracts`	Curated, publicly announced deals; recall-low	"Announced [date]: [vendor] contract, value $X as announced". $0 = undisclosed. No renewal odds
`company_fai`	Derived statistic from observed product and role signals	"FAI: Sales 58% / Engineering 8% of usage". No extra hedge needed
`company_intent` topics	Observed research activity, modeled score	"HG intent on [topic], score N". Only with resolved topics
TrustRadius activities (`company_intent` `activities`)	Observed first-party buyer activity on TrustRadius; no score	"Buyers at the company viewed [category] reviews on TrustRadius". Check relevance before calling it a buying signal
`get_product_reviews`	Observed reviewer opinions (self-reported)	"TrustRadius reviewers report..."
Contacts (`contact_search`, `contact_enrich`)	Third-party provider data, not HG	"Per third-party contact data". Titles can lag; verify before quoting
SEC filings	Observed, company-reported	"Per the FY2025 10-K"
Web search	Depends on the source; the `answer` field is an AI summary	"Per [domain.com]". Never cite the summary

Decision rule

Use the strongest qualifier the source supports. When in doubt, say "modeled" or "estimated". The reader will trust the brief more if the qualifier is upfront than if a CRO catches a precise-looking number that turns out to be an estimate.

Don't upgrade a value. A band stays a band, a derived label stays context, and placeholder values (2000-01-01, 1970-01-01, -1) are "no value", never a fact.

Keep each value tied to its tool. The same metric can differ slightly between tools in one session. Quote each value with the tool it came from and don't mix them in one claim.

When to drop the qualifier

The first mention of each source carries the qualifier; later mentions in the same section can drop it. "HG models the company's IT spend at $50M; security accounts for about 6% of that": the second clause drops "models" because the frame is set.

In a table sourced entirely from company_spend, one introductory sentence ("HG modeled spend, retrieved Apr 2026") covers every cell.

Wrong: a heading "HG modeled estimate", then "These numbers are HG model estimates", then a "Modeled Spend" column, then "on the model's numbers".
Right: one qualifier where the section starts ("HG modeled IT spend, Oct 2026"), then plain figures.

Common pitfalls

Saying "they spend $X" without a qualifier. It reads like an invoice extract.
Saying "modeled estimate of $X" twice in a row. Once is enough.
Calling contracts "modeled" or inferring renewal odds from them.
Calling a company an "AI leader" from an operating-signal label.
Calling intent "active research" without resolved topics or a recent latest_signal_date. Pair with hg-recent-signals.
Presenting contact data as HG data.

Reference

V2 tool docs: https://phoenix.hginsights.com/docs/mcp-tools/v2/overview
Full docs for LLMs: https://phoenix.hginsights.com/docs/llms-full.txt
Related skills: hg-mcp-calling-rules, hg-citation-discipline, hg-recent-signals

Skill: HG Modeled vs. Observed Data