---
title: "How evidence labels work: the four source labels explained"
description: "Evidence labels say where each claim came from: stated, worked out, corrected or added by your team, or inferred. What each one means, with a worked example."
image: https://www.marketingmary.ai/hubfs/blog/images/how-evidence-labels-work-hero.webp
---

[Skip to main content](https://www.marketingmary.ai/blog/how-evidence-labels-work#main-content)

[![](https://www.marketingmary.ai/hs-fs/hubfs/mm-2026-preview/marketingmary-logo-mark-only.png?width=259&height=155&name=marketingmary-logo-mark-only.png) Marketing Mary](https://www.marketingmary.ai/)

- [How it works ▾](https://www.marketingmary.ai/how-it-works) How it works
  
    - [OverviewThe whole journey, step by step](https://www.marketingmary.ai/how-it-works)
    - [PersonasProfiles of your best customers](https://www.marketingmary.ai/personas)
    - [ArticlesWritten, checked and approved](https://www.marketingmary.ai/content)
    - [The 90-day planComing later · week by week, sized to your hours](https://www.marketingmary.ai/90-day-plan)
    - [How Mary checks her workSources, labels and checks](https://www.marketingmary.ai/methodology)
- [Who it’s for ▾](https://www.marketingmary.ai/for-small-businesses) Who it’s for
  
    - [Small businessesOwners and one-person teams](https://www.marketingmary.ai/for-small-businesses)
    - [Marketing teamsPersonas and content your board can defend](https://www.marketingmary.ai/for-marketing-teams)
    - [AgenciesEvery client in one account](https://www.marketingmary.ai/for-agencies)
- [Pricing](https://www.marketingmary.ai/pricing)
- [Blog](https://www.marketingmary.ai/blog)

[Join the early-adopter list](https://www.marketingmary.ai/join-our-waitlist)

 9 min read

# How evidence labels work: stated, worked out, corrected or added by your team, inferred

[![Picture of Clwyd Probert](https://app.hubspot.com/settings/avatar/ce454bc741a43a0578a9342cd3f99842) Clwyd Probert](https://www.marketingmary.ai/blog/author/clwyd-probert) : Published November 21, 2026

[marketing](https://www.marketingmary.ai/blog/tag/marketing) [Buyer Personas](https://www.marketingmary.ai/blog/tag/buyer-personas)

![How evidence labels work: stated, worked out, corrected or added by your team, inferred](https://www.marketingmary.ai/hubfs/blog/images/how-evidence-labels-work-hero.webp)

**Evidence labels** are short tags that say where each claim about a company came from. Marketing Mary puts one of four on every statement: **stated on your website** (copied from a named page), **worked out from your website** (pieced together from several pages), **corrected or added by your team** (supplied or corrected by a person) or **inferred** (Mary's own reading, not stated anywhere).

A label doesn't tell you whether a claim is true. It tells you what the claim rests on, so you know which lines to trust, which to check and which to correct. Anything Mary couldn't check against its pages isn't shown as fact, and the screen says how many claims were left out. Below: what each label means, how quotes and corrections work, a worked example, and how to use the labels yourself.

Key takeaway

Each evidence label answers one question: where did this come from? **Stated** points at one sentence on one page. **Worked out** points at several pages. **Corrected or added by your team** points at a named person and a date. **Inferred** points at no one's words: it's Mary's reading, and it says so. A claim with nothing behind it isn't shown at all.

## What are evidence labels?

Evidence labels are source tags attached to individual claims rather than to a whole document. In Marketing Mary, every line of a company reading carries its own label. "Based in London", copied from an About page, and "sells mainly to scale-ups", pieced together from a pricing page and case studies, are different kinds of claim, and they look different on the screen.

Evidence labels are not the same as AI-disclosure labels. [C2PA Content Credentials](https://c2pa.org/), an open technical standard for establishing "the origin and edits of digital content", and the transparency rules in [Article 50 of the EU AI Act](https://artificialintelligenceact.eu/article/50/) deal with how a piece of content was made, such as whether a model generated or altered it. Evidence labels deal with where a single statement came from.

## Why does every claim need a label?

Every claim needs a label because fluent text and supported text look the same on the page. The US National Institute of Standards and Technology names the risk in its [Generative AI Profile](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf): "confabulation", or "the production of confidently stated but erroneous or false content". A link beside a claim doesn't settle the question either. Two studies show the gap:

11.0%

Claims not supported by their own citations

of 98,020 single claims in Google AI Overview answers were not supported by the pages those answers cited (55,393 queries, Mar 13 to Apr 21, 2026)

researched · Xu, Iqbal & Montgomery, 2026

4,927

US adults in a trust experiment

Reference links raised trust in AI search answers, even when the links were incorrect or hallucinated (preregistered experiment, run Mar to Apr 2024)

researched · Li & Aral, 2025

The first figure comes from [a 2026 measurement study of Google AI Overviews](https://arxiv.org/abs/2605.14021). Omission was the most common failure: no cited page mentioned the claim at all. The authors note that pages they couldn't read may inflate the figure; on their most generous assumption, about 5.3% of claims stay unsupported. The second comes from [Human trust in AI search](https://arxiv.org/abs/2504.06435), which found links raise trust "even when those links and citations are incorrect or hallucinated." A source has to be opened, and the claim checked against it.

Intelligence analysts have written this rule into their standards for years. The US Office of the Director of National Intelligence's [Intelligence Community Directive 203](https://archive.dni.gov/files/documents/ICD/ICD-203.pdf) asks analysts to "clearly distinguish statements that convey underlying intelligence information used in analysis from statements that convey assumptions or judgments." Evidence labels make that distinction for a company reading, one line at a time.

## What do the four evidence labels mean?

Each of Marketing Mary's four evidence labels points at a different kind of source. A fifth state, "couldn't be verified", isn't a label at all: it's a count of what was left out.

| Label | Where it came from | What you can open | How to use it |
| --- | --- | --- | --- |
| stated on your website | Copied from a named page on the website. | The exact sentence and the page it came from. | Lean on it as far as your website is current. |
| worked out from your website | Pieced together from what several pages say. | The list of pages it was worked out from. | Check the pages listed; the conclusion is Mary's, the pages are yours. |
| corrected or added by your team | Supplied or corrected by a person on the account. | The person's name, the date, and Mary's original reading, kept. | It's as good as the person's knowledge. |
| inferred | Mary's own reading. It isn't stated on any page. | The label itself, so you can tell it apart from what the website says. | Treat it as a judgment to confirm, and correct it if it's wrong. |
| couldn't be verified | A claim Mary couldn't check against the pages it came from. | Nothing: it isn't shown as fact. The screen counts it, for example "3 claims couldn't be verified and aren't shown". | There's nothing to lean on, and the count tells you so. |

The four labels and the "couldn't be verified" count, as Marketing Mary uses them. Definitions follow the [How Mary checks her work](https://www.marketingmary.ai/methodology) page.

The four labels line up with an existing web standard. The W3C's [PROV ontology](https://www.w3.org/TR/prov-o/) separates a quotation ("the repeat of (some or all of) an entity"), a derivation ("a transformation of one entity into another"), a revision ("a revised version of some original") and attribution ("the ascribing of an entity to an agent"). Stated is a quotation, worked out is a derivation, and corrected or added by your team is a revised version with a named person attached. Inferred is what remains: Mary's own judgment, marked as hers.

![Marketing Mary company overview for a fictional company, Brightmoor. Name, industry and location are labeled stated on your website; markets served and target markets are labeled worked out from your website; size signals are labeled corrected or added by your team. Two open labels show the quoted sentence with its About page, and the pricing page a target market was worked out from. A note says 3 things couldn't be verified and aren't shown](https://www.marketingmary.ai/hs-fs/hubfs/mm-2026/crops/for-marketing-teams-02-overview-quote.webp?width=2400&height=821&name=for-marketing-teams-02-overview-quote.webp)

Marketing Mary's company overview: every line carries its label, and opening one shows the sentence and page, or the pages it was worked out from. Real screen with sample data; Brightmoor is a fictional company.

## How does a quote get from your website to a "stated" label?

![Ink illustration of tweezers lifting a single printed line out of a page and placing it unchanged on a card, with a capped pen set aside, unused](https://www.marketingmary.ai/hs-fs/hubfs/blog/images/quote-extracted-not-written.webp?width=600&height=600&name=quote-extracted-not-written.webp)

"Stated on your website" is the strictest evidence label, because it carries a quote. Marketing Mary handles quotes in one specific way: Mary isn't allowed to write one. When Mary makes a claim, she names the pages that support it. Marketing Mary then pulls the exact sentence from its stored copy of that page and checks that it's there.

If the check passes, you see the claim with its quote and its page. If it fails, the claim is left out, and the screen says how many were: "3 claims couldn't be verified and aren't shown."

This matters because an invented quote is the most convincing kind of error. Because quotes are pulled from the stored page rather than written by a model, a made-up quotation is prevented by design, not just discouraged. The full sequence is on the [How Mary checks her work](https://www.marketingmary.ai/methodology) page.

## What's the difference between "worked out from your website" and "inferred"?

Both labels mark a conclusion rather than a quote. The difference is whether you can see the pages behind it. worked out from your website means Mary pieced the claim together from what several pages say and lists those pages, for example "mainly across the UK and Ireland", worked out from nine pages of a sample company's site. inferred means the claim isn't stated on any page. It's Mary's own reading, such as an estimate of a company's size, and it's always labeled that way.

Inferred isn't a polite word for wrong. Plenty of useful conclusions aren't written down anywhere. The label tells you the line is a judgment, so it's the first place to check and the easiest to correct.

Buyer personas use a related scale. Each persona shows its evidence mix: how much rests on first-party evidence (the company's own website), on researched evidence (outside research, with sources) and on synthesis (Mary's reasoning, labeled as such), with a confidence level of high, medium or low. Our guide to the [evidence-graded buyer persona](https://www.marketingmary.ai/blog/evidence-graded-buyer-persona) explains that scale, and the [personas page](https://www.marketingmary.ai/personas) shows it on a dossier.

## What happens when you correct a claim?

When you correct a claim in Marketing Mary, your correction is marked corrected or added by your team, with your name and the date. Mary's original reading is kept, so anyone can see what changed, and you can go back to it at any time. The correction also carries forward: fix who the company sells to, and the next set of personas is built on what you said.

![A corrected fact in Marketing Mary: size signals now read 38 people, marked as corrected by a named person on Sep 3, 2026. A panel shows the analysis originally inferred around 35 people, with options to edit, revert to the analysis or see the history (fictional sample data)](https://www.marketingmary.ai/hs-fs/hubfs/mm-2026/crops/methodology-06-correction.webp?width=2118&height=808&name=methodology-06-correction.webp)

A correction: Mary's inferred reading ("around 35 people") is kept beside the person's correction, with their name and the date. Real screen with sample data.

## What do evidence labels look like on a real company reading?

Here is how the four evidence labels look on one sample company. Harbour Supply is a fictional industrial supplier, the sample company used across this site. Every line below is illustrative sample data, not a real analysis.

Harbour Supply · company reading

harboursupply.com.au · 6 of 7 pages read · 1 page couldn't be read: /en-au/careers

What they doParts and supplies for mining and construction sitesstated on your website · Home page

Delivery promiseNext-day deliverystated on your website · Home page

Who they sell toProcurement teams at mining and construction firmsworked out from your website · 4 pages

Markets servedMining and construction firms, plus local councilscorrected or added by your team · Sam Rivera · Oct 14, 2026

SizeA small team on one siteinferred · Mary's reading

Mary's original reading of markets served, kept: "Mostly mining companies" (inferred) · 1 claim couldn't be verified and isn't shown

Illustrative card with sample data. Harbour Supply and Sam Rivera are fictional.

The two stated lines you can confirm by opening the Home page. The worked-out line names its four pages, so you can check the reasoning. The size line is a judgment, so it's the one to confirm first. And the correction shows what "corrected or added by your team" is for: Mary's reading of the market was narrower than the business, the owner fixed it, and the next personas start from the corrected market.

The same labels carry through to personas. On the sample persona Procurement Priya, "buys parts for mining and construction sites" comes from the Home page, "worries about a job stalling while she waits for parts" is worked out from four pages, and "says yes when opening a trade account is quick" is inferred by Mary. Her card shows medium confidence and an evidence mix of 18% first-party, 64% researched and 18% synthesis.

## How can you use evidence labels in your own research?

You can apply the same four evidence labels to a persona, a content brief or a competitor summary written by hand. A spreadsheet is enough to start:

1. **Split the document into single claims.** One row per statement. "A fast-growing, London-based firm that sells to scale-ups" is three claims, not one.
2. **Give each claim one label and keep its evidence beside it.** The sentence and the page for stated; the list of pages for worked out; the person and the date for corrected or added by your team. Inferred needs no source, which is the point of the label.
3. **Use quotation marks only for copied words.** A paraphrase is worked out, not stated. If you can't point at the sentence, don't quote it.
4. **Keep the original when you correct something.** Strike it through rather than deleting it, and add your name and the date.
5. **Count what you left out.** "2 claims couldn't be checked" is information. A blank is better than a confident guess.
6. **Date the reading.** Websites change. A label describes the evidence on the day you looked.

That's what Marketing Mary does on every analysis: you give Mary a web address, she reads the site, and every line comes back labeled. It's also the rule we apply to articles, set out in our guide to [what to automate in content creation, and where a person signs off](https://www.marketingmary.ai/blog/automated-content-creation-pipeline): take quotes from the source, never from the model.

## Frequently asked questions

### What does "inferred" mean in Marketing Mary?

"Inferred" marks Mary's own reading of a company: something that isn't stated on any page of its website. It's always labeled, so you can tell it apart from what the website actually says, and you can correct it if it's wrong.

### Is "worked out from your website" less reliable than "stated on your website"?

It's a different kind of evidence. A stated claim points at one sentence on one page. A worked-out claim is pieced together from several pages, and Marketing Mary lists those pages, so you can check the reasoning yourself.

### Can Marketing Mary make up a quote?

No. Mary names the pages that support a claim, and Marketing Mary pulls the quote from its stored copy of that page and checks it. If the check fails, the claim isn't shown, and the screen counts how many were left out.

### What happens to Mary's original reading when I correct a claim?

It's kept. Your correction is marked "corrected or added by your team" with your name and the date, and Mary's original reading stays alongside it. You can go back to it at any time, and what Mary builds next follows your correction.

### Are evidence labels the same as labels for AI-generated content?

No. AI-content labels, such as C2PA Content Credentials or the transparency rules in Article 50 of the EU AI Act, describe how content was made. Evidence labels describe where a single claim came from, and they apply to research written by hand too.

See what Mary finds on your website.

Join the early-adopter list and be one of the first to try Mary on your own website.

[Join the early-adopter list](https://www.marketingmary.ai/join-our-waitlist)

Plans from Free to Enterprise · pricing to be confirmed

Clwyd Probert

Founder, Marketing Mary

Clwyd Probert is the founder of Marketing Mary, an evidence-first persona intelligence platform for B2B marketing, and CEO of Whitehat, a London-based SEO and inbound marketing agency and HubSpot Diamond Solutions Partner.

Sources, checked against the original documents on Nov 1, 2026

researched  [Xu, Iqbal and Montgomery, Measuring Google AI Overviews: activation, source quality, claim fidelity, and publisher impact](https://arxiv.org/abs/2605.14021). arXiv 2605.14021, v2, Oct 1, 2026.

researched  [Li and Aral, Human trust in AI search: a large-scale experiment](https://arxiv.org/abs/2504.06435). arXiv 2504.06435, Apr 2025.

researched  [W3C, PROV-O: The PROV Ontology](https://www.w3.org/TR/prov-o/). W3C Recommendation, Apr 30, 2013.

researched  [NIST AI 600-1, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf). NIST, Jul 2024.

researched  [Intelligence Community Directive 203: Analytic Standards](https://archive.dni.gov/files/documents/ICD/ICD-203.pdf). ODNI, 2015, as amended.

researched  [Coalition for Content Provenance and Authenticity (C2PA), Content Credentials](https://c2pa.org/). Accessed Nov 1, 2026.

researched  [EU AI Act, Article 50: Transparency obligations for providers and deployers of certain AI systems](https://artificialintelligenceact.eu/article/50/). AI Act Explorer, accessed Nov 1, 2026.

first-party  How Marketing Mary labels claims, extracts quotes and keeps corrections: Marketing Mary's own product documentation. Screens and cards show fictional sample data.

- [Tweet](https://twitter.com/share)

[![How evidence labels work: stated, worked out, corrected or added by your team, inferred](https://www.marketingmary.ai/hubfs/blog/images/how-evidence-labels-work-hero.webp)](https://www.marketingmary.ai/blog/how-evidence-labels-work)

#### [How evidence labels work: stated, worked out, corrected or added by your team, inferred](https://www.marketingmary.ai/blog/how-evidence-labels-work)

[![Picture of Clwyd Probert](https://app.hubspot.com/settings/avatar/ce454bc741a43a0578a9342cd3f99842) Clwyd Probert](https://www.marketingmary.ai/blog/author/clwyd-probert): Nov 21, 2026, 9:16:00 AM

Evidence labels are short tags that say where each claim about a company came from. Marketing Mary puts one of four on every statement: stated on...

[marketing](https://www.marketingmary.ai/blog/tag/marketing) [Buyer Personas](https://www.marketingmary.ai/blog/tag/buyer-personas)

[Read More](https://www.marketingmary.ai/blog/how-evidence-labels-work)

[![What is an evidence-graded buyer persona?](https://www.marketingmary.ai/hubfs/blog/images/evidence-graded-buyer-persona-hero.webp)](https://www.marketingmary.ai/blog/evidence-graded-buyer-persona)

#### [What is an evidence-graded buyer persona?](https://www.marketingmary.ai/blog/evidence-graded-buyer-persona)

[![Picture of Clwyd Probert](https://app.hubspot.com/settings/avatar/ce454bc741a43a0578a9342cd3f99842) Clwyd Probert](https://www.marketingmary.ai/blog/author/clwyd-probert): Sep 30, 2026, 8:25:00 PM

An evidence-graded buyer persona is a buyer persona in which every claim carries a label saying where it came from: first-party evidence (the...

[marketing](https://www.marketingmary.ai/blog/tag/marketing) [Buyer Personas](https://www.marketingmary.ai/blog/tag/buyer-personas)

[Read More](https://www.marketingmary.ai/blog/evidence-graded-buyer-persona)

[![Impressions Up, Clicks Down: What's Really Happening in Search Console (2026)](https://www.marketingmary.ai/hubfs/blog/images/aeo-06-impressions-clicks-divergence-her-hero-1.webp)](https://www.marketingmary.ai/blog/impressions-up-clicks-down)

#### [Impressions Up, Clicks Down: What's Really Happening in Search Console (2026)](https://www.marketingmary.ai/blog/impressions-up-clicks-down)

[![Picture of Clwyd Probert](https://app.hubspot.com/settings/avatar/ce454bc741a43a0578a9342cd3f99842) Clwyd Probert](https://www.marketingmary.ai/blog/author/clwyd-probert): Sep 9, 2026, 11:58:11 PM

Our own Search Console tells this story: impressions up 53% and clicks down 6% in the 28 days to September 6, 2026, compared with the 28 days before,...

[Read More](https://www.marketingmary.ai/blog/impressions-up-clicks-down)

[![Marketing Workflow Automation: 5 Workflows for B2B Teams](https://www.marketingmary.ai/hubfs/blog/images/marketing-workflow-automation-hero-hero-1.webp)](https://www.marketingmary.ai/blog/marketing-workflow-automation-guide)

#### [Marketing Workflow Automation: 5 Workflows for B2B Teams](https://www.marketingmary.ai/blog/marketing-workflow-automation-guide)

[![Picture of Clwyd Probert](https://app.hubspot.com/settings/avatar/ce454bc741a43a0578a9342cd3f99842) Clwyd Probert](https://www.marketingmary.ai/blog/author/clwyd-probert) : Mar 23, 2026, 12:00:00 PM

Marketing teams lose hours every week to repetitive tasks. See five AI marketing workflows, from brief to reporting, and what to automate in each.

[AI](https://www.marketingmary.ai/blog/tag/ai) [marketing](https://www.marketingmary.ai/blog/tag/marketing) [Marketing Automation](https://www.marketingmary.ai/blog/tag/marketing-automation)

[Read More](https://www.marketingmary.ai/blog/marketing-workflow-automation-guide)

[![What is an evidence-graded buyer persona?](https://www.marketingmary.ai/hubfs/blog/images/evidence-graded-buyer-persona-hero.webp)](https://www.marketingmary.ai/blog/evidence-graded-buyer-persona)

#### [What is an evidence-graded buyer persona?](https://www.marketingmary.ai/blog/evidence-graded-buyer-persona)

[![Picture of Clwyd Probert](https://app.hubspot.com/settings/avatar/ce454bc741a43a0578a9342cd3f99842) Clwyd Probert](https://www.marketingmary.ai/blog/author/clwyd-probert) : Sep 30, 2026, 8:25:00 PM

An evidence-graded buyer persona labels every claim by source: first-party, researched or synthesis. What it means, why it matters and how to grade...

[marketing](https://www.marketingmary.ai/blog/tag/marketing) [Buyer Personas](https://www.marketingmary.ai/blog/tag/buyer-personas)

[Read More](https://www.marketingmary.ai/blog/evidence-graded-buyer-persona)

[![Automated content creation: what to automate, and where a person signs off](https://www.marketingmary.ai/hubfs/blog/images/automated-content-creation-signoff-hero--hero-1.webp)](https://www.marketingmary.ai/blog/automated-content-creation-pipeline)

#### [Automated content creation: what to automate, and where a person signs off](https://www.marketingmary.ai/blog/automated-content-creation-pipeline)

[![Picture of Clwyd Probert](https://app.hubspot.com/settings/avatar/ce454bc741a43a0578a9342cd3f99842) Clwyd Probert](https://www.marketingmary.ai/blog/author/clwyd-probert) : Mar 24, 2026, 9:00:00 AM

What an automated content pipeline should do from brief to publishing, which steps are safe to automate, and where a person must check and approve...

[AI](https://www.marketingmary.ai/blog/tag/ai) [marketing](https://www.marketingmary.ai/blog/tag/marketing)

[Read More](https://www.marketingmary.ai/blog/automated-content-creation-pipeline)

![](https://www.marketingmary.ai/hs-fs/hubfs/mm-2026-preview/marketingmary-logo-mark-only.png?width=259&height=155&name=marketingmary-logo-mark-only.png) Marketing Mary

How it works

- [Overview](https://www.marketingmary.ai/how-it-works)
- [Personas](https://www.marketingmary.ai/personas)
- [Articles](https://www.marketingmary.ai/content)
- [The 90-day plan](https://www.marketingmary.ai/90-day-plan)
- [How Mary checks her work](https://www.marketingmary.ai/methodology)
- [Pricing](https://www.marketingmary.ai/pricing)
- [Managed](https://www.marketingmary.ai/managed)

Who it’s for

- [Small businesses](https://www.marketingmary.ai/for-small-businesses)
- [Marketing teams](https://www.marketingmary.ai/for-marketing-teams)
- [Agencies](https://www.marketingmary.ai/for-agencies)

Company

- [Blog](https://www.marketingmary.ai/blog)
- [Security](https://www.marketingmary.ai/security)
- [Privacy](https://www.marketingmary.ai/privacy-policy)
- [Terms](https://www.marketingmary.ai/terms-of-service)
- [Contact](https://www.marketingmary.ai/contact)

© 2026 Marketing Mary

```json
{
  "@context" : "https://schema.org",
  "@type" : "BlogPosting",
  "author" : {
    "@type" : "Person",
    "name" : "Clwyd Probert",
    "url" : "https://www.marketingmary.ai/blog/author/clwyd-probert"
  },
  "datePublished" : "2026-11-21T09:16:00.000Z",
  "headline" : "How evidence labels work: the four source labels explained",
  "image" : [ "https://www.marketingmary.ai/hubfs/blog/images/how-evidence-labels-work-hero.webp" ],
  "mainEntityOfPage" : {
    "@id" : "https://www.marketingmary.ai/blog/how-evidence-labels-work",
    "@type" : "WebPage"
  },
  "publisher" : {
    "@type" : "Organization",
    "logo" : {
      "@type" : "ImageObject",
      "url" : "https://www.marketingmary.ai/hubfs/Logo%20White-1.png"
    },
    "name" : "Marketing Mary"
  }
}
```

```json
{
  "@context" : "https://schema.org",
  "@type" : "Blog",
  "description" : "Plain-English guides to marketing plans, buyer personas, content and getting found on Google and in AI answers, from the team behind Marketing Mary.",
  "inLanguage" : "en-US",
  "name" : "Marketing Mary Blog",
  "publisher" : {
    "@type" : "Organization",
    "name" : "Marketing Mary",
    "url" : "https://www.marketingmary.ai"
  },
  "url" : "https://www.marketingmary.ai/blog"
}
```

```json
{
  "@context" : "https://schema.org",
  "@id" : "https://www.marketingmary.ai/blog/how-evidence-labels-work#article",
  "@type" : "BlogPosting",
  "about" : {
    "@type" : "Thing",
    "name" : "Evidence labels"
  },
  "author" : {
    "@type" : "Person",
    "jobTitle" : "Founder, Marketing Mary",
    "name" : "Clwyd Probert",
    "url" : "https://www.marketingmary.ai/blog/author/clwyd-probert",
    "worksFor" : {
      "@id" : "https://www.marketingmary.ai/#organization"
    }
  },
  "dateModified" : "2026-11-21",
  "datePublished" : "2026-11-21",
  "description" : "Evidence labels say where each claim came from: stated, worked out, corrected or added by your team, or inferred. What each one means, with a worked example.",
  "headline" : "How evidence labels work: stated, worked out, corrected or added by your team, inferred",
  "image" : "https://27046115.fs1.hubspotusercontent-eu1.net/hubfs/27046115/blog/images/how-evidence-labels-work-hero.webp",
  "inLanguage" : "en-US",
  "keywords" : [ "evidence labels", "source labels", "how to label sources", "stated on your website", "worked out from your website", "corrected or added by your team", "inferred", "claim-level provenance" ],
  "mainEntityOfPage" : {
    "@id" : "https://www.marketingmary.ai/blog/how-evidence-labels-work",
    "@type" : "WebPage"
  },
  "publisher" : {
    "@id" : "https://www.marketingmary.ai/#organization"
  },
  "wordCount" : 2516
}
```

```json
{
  "@context" : "https://schema.org",
  "@id" : "https://www.marketingmary.ai/#organization",
  "@type" : "Organization",
  "logo" : {
    "@type" : "ImageObject",
    "url" : "https://www.marketingmary.ai/hubfs/Marketingmary%20logo-02.png"
  },
  "name" : "Marketing Mary",
  "url" : "https://www.marketingmary.ai/"
}
```

```json
{
  "@context" : "https://schema.org",
  "@id" : "https://www.marketingmary.ai/blog/how-evidence-labels-work#faq",
  "@type" : "FAQPage",
  "mainEntity" : [ {
    "@type" : "Question",
    "acceptedAnswer" : {
      "@type" : "Answer",
      "text" : "\"Inferred\" marks Mary's own reading of a company: something that isn't stated on any page of its website. It's always labeled, so you can tell it apart from what the website actually says, and you can correct it if it's wrong."
    },
    "name" : "What does \"inferred\" mean in Marketing Mary?"
  }, {
    "@type" : "Question",
    "acceptedAnswer" : {
      "@type" : "Answer",
      "text" : "It's a different kind of evidence. A stated claim points at one sentence on one page. A worked-out claim is pieced together from several pages, and Marketing Mary lists those pages, so you can check the reasoning yourself."
    },
    "name" : "Is \"worked out from your website\" less reliable than \"stated on your website\"?"
  }, {
    "@type" : "Question",
    "acceptedAnswer" : {
      "@type" : "Answer",
      "text" : "No. Mary names the pages that support a claim, and Marketing Mary pulls the quote from its stored copy of that page and checks it. If the check fails, the claim isn't shown, and the screen counts how many were left out."
    },
    "name" : "Can Marketing Mary make up a quote?"
  }, {
    "@type" : "Question",
    "acceptedAnswer" : {
      "@type" : "Answer",
      "text" : "It's kept. Your correction is marked \"corrected or added by your team\" with your name and the date, and Mary's original reading stays alongside it. You can go back to it at any time, and what Mary builds next follows your correction."
    },
    "name" : "What happens to Mary's original reading when I correct a claim?"
  }, {
    "@type" : "Question",
    "acceptedAnswer" : {
      "@type" : "Answer",
      "text" : "No. AI-content labels, such as C2PA Content Credentials or the transparency rules in Article 50 of the EU AI Act, describe how content was made. Evidence labels describe where a single claim came from, and they apply to research written by hand too."
    },
    "name" : "Are evidence labels the same as labels for AI-generated content?"
  } ]
}
```

```json
{
  "@context" : "https://schema.org",
  "@id" : "https://www.marketingmary.ai/blog/how-evidence-labels-work#breadcrumb",
  "@type" : "BreadcrumbList",
  "itemListElement" : [ {
    "@type" : "ListItem",
    "item" : "https://www.marketingmary.ai/",
    "name" : "Home",
    "position" : 1
  }, {
    "@type" : "ListItem",
    "item" : "https://www.marketingmary.ai/blog",
    "name" : "Blog",
    "position" : 2
  }, {
    "@type" : "ListItem",
    "item" : "https://www.marketingmary.ai/blog/how-evidence-labels-work",
    "name" : "How evidence labels work",
    "position" : 3
  } ]
}
```