First-Party Evidence Blocks Editors Must Keep When AI Drafts

By Simeon Matheka, Founder & Creative Director · Published 2026-10-01 · Updated 2026-10-01 · 12 min read

The first thing a model does to a page is smooth it. Screenshots become “in our experience,” dated tests lose their dates, and the one detail only you had disappears. How to mark, protect, and verify the evidence that makes the page worth publishing.

You ask a model to tighten a section. It comes back clean, shorter, and a little smarter-sounding. It also lost the date of your test, the odd failure that made the section worth reading, and the name of the tool you actually used. Nobody noticed, because the new version reads better.

This is the editor’s job in an AI-assisted workflow: keep the parts only you could have written. It is the second step after the review gates, which ask whether a first-party block exists. This article covers what counts, how to mark it, and how to prove it survived. The printable version is the evidence block checklist.

Why evidence blocks are the page

Sourced from Google. Its helpful-content self-assessment asks whether content provides original information, reporting, research, or analysis, and whether it clearly demonstrates first-hand expertise and depth of knowledge, for example expertise that comes from having used a product or service. Source: Creating helpful, reliable, people-first content (Search Central, checked 1 October 2026).

The generative AI optimization guide goes further. It says unique, compelling, useful content will likely influence your presence in generative AI search in the long run more than any other suggestion in that guide. It contrasts commodity content (common knowledge from anyone) with non-commodity content that gives unique expert or experienced takes, and tells you not to recycle what others have said or what a generative AI model could easily produce. Source: Optimizing your website for generative AI features on Google Search (Search Central, checked 1 October 2026).

Put those together and the editorial rule is plain. A model can supply structure and tidy prose. It cannot supply your evidence. If the evidence goes, so does most of the reason to publish.

What counts as a first-party block

Not every “we” sentence is evidence. “In our experience, speed matters” is a claim with a costume on. An evidence block is something a reader or reviewer could check or at least test against their own situation. The labels below match the ones we use on every article here.

The fifth row is the dangerous one. Never ship an invented client, percentage, or testimonial. If you do not have a real, permissioned fact, teach the method instead and say plainly that the example is constructed.

Mark blocks before the model touches the draft

The cheapest protection is a visible marker around every evidence block in your working document. You are not building software. You are giving the editor and the model a clear “do not rewrite this” signal and giving yourself something to search for later.

[KEEP: observed | owner: Sam | date: 2026-09-30]
We ran the form on a staging copy with JavaScript disabled.
The submit button did nothing. The fix was a plain HTML form action.
Does not prove: how this behaves for every browser.
[/KEEP]

[KEEP: sourced | link: https://developers.google.com/search/docs/crawling-indexing/links-crawlable]
Google can generally only crawl a link that is an <a> element with an href attribute.
[/KEEP]

Instruction to the model: you may reformat or fix grammar inside KEEP blocks.
You may not change numbers, dates, names, tool names, or the sentence
that starts "Does not prove". If a KEEP block seems wrong, leave it and
add a comment after it.

This is a framework, not a guarantee. Models do not always obey instructions about what to leave alone, which is why the diff check below exists. Treat the marker as a first line of defense and the diff as the real one.

The diff check editors actually run

1. Before the AI pass, copy the text of every KEEP block into a separate note, or save the version.
2. After the pass, search the new draft for each block. Compare it word by word, not by feel. Use a text compare tool if you have one.
3. Check six things inside each block: numbers, dates, names, tool names, links, and the “does not prove” line. If any changed, restore the original.
4. Check what moved. A block that is now in a different section may have lost the context that made it true. Put it back near the claim it supports.
5. Remove the markers before publish. They are working notes, not copy. Keep the underlying labels in the visible text (“Observed on 30 September 2026”).

Where the model is allowed to help

AI is useful here, within limits. Google’s generative AI guidance says the tools can be useful when researching a topic and for adding structure to original content. Source: Google Search’s guidance on using generative AI content on your website (Search Central, checked 1 October 2026). In practice, that gives you a clean division of labor.

Filled hypothetical: one page, three blocks

Hypothetical example, labeled as such. A studio drafts a page on “why our contact form stopped working after a redesign,” with a model assisting.

Without the diff, the second block would have shipped as a vague claim. With it, the editor restored two details in two minutes. That is the whole process. It is boring and it works because it is boring.

Failure modes

• Evidence after the fact: someone adds an “in our experience” paragraph at the end to pass the review. Fix: gather evidence first, then draft around it.
• Unlabeled hypotheticals: a constructed example drifts into reading as a real client story. Fix: say “Hypothetical” in the sentence, every time.
• Screenshot with no caption: readers cannot tell what to notice. Fix: one sentence on what the image proves and what it does not.
• Private data in a proof block: customer names, emails, or account IDs in screenshots. Fix: redact before upload, or do not use it.
• Block buried at the bottom: the best material hides in a closing paragraph. Fix: lead the section with it.

Ship checks

1. The page has at least one evidence block, and its label is visible in the text, not only in your notes.
2. The six details (number, date, name, tool, link, limit) match the saved originals.
3. No invented client, percentage, quote, or lift figure appears anywhere, including the meta description and FAQs.
4. Each image has descriptive alt text and a caption saying what it shows.
5. The owner who ran the test or took the screenshot has read the final page.

Print the evidence block checklist and run it on your next AI-assisted draft. For the broader argument about why added material matters, read the information gain framework. If you are rebuilding a site so that proof like this has a home, get in touch.

Tags: first-party evidence, AI content, editorial standards, information gain, E-E-A-T, content operations