How to Harvest, Inventory and Repurpose Everything the Market Already Says About You

Most entrepreneurs are sitting on years of proof they have never collected. Here is the exact process for finding it, verifying it, scoring it, and putting it on the site — plus the four traps that make the difference between an audit and a result.

This is a field-run supporting SOP; the definitive task is linked above. It was written from a live run on a client site in August 2026, and every number in the worked example is real. Skills referenced: positive-mentions-harvester, proof-to-page-repurposing, archive-attribution-defense, geo-visibility-audit, evidence-verification.

The short version

1. Lock the canonical facts first. One agreed version of every title, date and number, in writing, from the client.
2. Harvest everything the market already said. Every podcast, press mention, award, speaking listing, testimonial and unprompted shoutout — each with a URL you actually fetched.
3. Score it on 30 points. Source authority, specificity, buy-box relevance.
4. Check where each item currently lives, and how many links point at that page.
5. Give every item a destination and a second use.
6. Publish the two pages nobody has: a proof page, and a canonical record of the flagship credential.
7. Ask the domains that already write about them for a link.
8. Re-verify against the live site before you report anything.

1. Canonical facts before anything else

You cannot harvest proof for a person whose own pages disagree with each other. Before searching, get one written answer for each of: current title, previous titles and their exact dates, every revenue or scale number, awards, education, and the count of anything countable.

Write it as a table with a fourth column headed “never write.” That column is the one that saves you. In the August run it held: never write “$45M at IPO listing” (the company never listed), never write “165+ keynotes” (they were partner trainings), never write “Chief Commercial Officer” (a title that appears in an AI encyclopedia and in no primary source).

Every number on a bio is a claim. A claim with a source behind it is an asset. A claim without one is a liability that sits quietly until the wrong person checks it.

2. Harvest — and search the way the person actually lived

Search their name against every client, event, employer and topic. Then search the things they made, separately from the person. Then search in every language they have worked in — an English-only search of a non-English career measures your search, not their authority.

Angles that reliably produce items nobody remembers:

• Former employer newsrooms — promotion and appointment releases often quote a CEO praising them by name, on a high-authority domain, and are used by nobody.
• Trade press for their industry, not general business press.
• Podcasts where they were the guest. Apple, Spotify, Podchaser and Listen Notes all index differently.
• Conference speaker pages and association member profiles.
• Critical articles. A former colleague writing a deliberately unflattering review of your client’s old employer, who then praises your client by name, is stronger proof than any testimonial — precisely because the piece was not trying to help.
• AI encyclopedias. They frequently hold a fuller biography than anything the client owns, along with a reference list that is a ready-made source hunt.

3. Score on 30 points

24–30, lead with it. 15–23, supporting proof. Under 15, archive it — weak proof dilutes strong proof. Score against the buy box, not the ego: a flattering mention that attracts the wrong work scores low.

And keep a separate gaps list. A claim you have not been able to source is a gap, not a low score. “We have not looked yet” must never render as “they scored badly.”

4. The step everyone skips: where does the proof currently live?

For every page that already holds proof, count the internal links pointing at it. This takes minutes and it is where the real finding usually is.

In the August run, the client had 43 verified proof items. Almost none were on his own site. The single client testimonial that was sat on a page with two inbound links that was not in the navigation — alongside every dated speaking engagement, an industry award, and four conference credits. A photo gallery holding 101 images had zero inbound links from anywhere, including the nav.

Nothing was missing. It was unreachable. Proof on a page nobody can reach is not proof.

5. Give every item a destination and a second use

Two columns. Which page it lands on, and what else it becomes — a quote card, a schema property, an encyclopedia citation, a $1/day creative. An item with no destination is not proof, it is trivia.

6. The two pages nobody has

A proof page. Every verified mention, ordered by score, each with a link that resolves. This is the page a programme committee reads, and the page an AI crawler reads, and almost no personal brand site has one.

A canonical record of the flagship credential. If the biggest thing on someone’s CV lives on a platform they do not control, check what that platform currently says about it. In the August run, a client had created a podcast and hosted 269 episodes over six years. The show descriptions on Apple, Spotify and RedCircle now name a different host. His name survives only in the URL slugs.

Meanwhile that same show was the number-one answer ChatGPT returned for a category query — a number-one ranking that named somebody else. That is worse than no ranking, because it proves the audience exists and is being handed to another name.

The fix has two halves and teams keep choosing one. Publish the canonical record yourself immediately, because it is entirely within your control. And pursue the upstream correction in parallel, because six podcast directories reading one RSS feed is one edit, not six conversations.

7. Ask for the link

List the domains that get quoted as authority about this person — by journalists, by AI engines, by anyone. Then check which of them link to the person’s site.

The usual answer is none. In the August run, seven employer and industry domains were being cited by all four AI engines as authority on the client, and not one of them linked to his site. His own pages carried the corrected facts and the models had no path to prefer them, so a stale sentence in an old press release kept winning.

Those seven domains already mention him by name. Nobody had ever asked them for a link.

The four traps

Trap 1: proof the models cannot read

Fetch every source URL twice — once with a browser user agent, once with a plain crawler agent. Two of the strongest items in the August library returned 403 to crawlers and rendered perfectly in a browser. They convince a human and are invisible to every AI model. They still belong in the library; they must be labelled, and the publisher asked to allow crawlers on that one page. That is a smaller ask than a new placement and it converts proof you already own.

Trap 2: proof rots

In one month, that client lost an award page to a publisher rebrand, an employer press release to a deletion, and a podcast host’s episode page to a 404 that still ranks first in Google. Archive every top-tier source to a citable location the day you score it, and re-check the whole library on every rerun — not just the new rows.

Trap 3: never let a matcher write a canonical tag

Sites built by repurposing an archive carry everything twice — the episode stub and the article written from it, both indexable, neither canonicalised. In the August run that was the direct cause of 27 of 58 articles missing from Google’s index.

The fix is obvious and the method is not. Exact title matching found 1 of 41 pairs, because platforms truncate titles and articles get retitled. Loosening to token-overlap scoring found 41 — and four of them were wrong: two different podcast appearances sharing only the client’s name, an article about one guest matched to an episode about another. A wrong canonical does not fail loudly. It quietly removes a page that deserved to rank, and nobody notices for months.

Use confidence scores to order a human’s reading list, not to replace it. Forty-five pairs took a few minutes to read.

Trap 4: never carry a finding forward

Before shipping the August repair plan, every finding from an audit two days old was re-tested against the live site. Six of thirteen had already been fixed. Shipping that list would have told a client we were about to fix six things that were already done.

The reverse error was worse: three items had been reported fixed and were not, and every one of them had a screenshot of the admin settings screen attached. A screenshot of a settings panel is not verification. Verify on the public surface the crawler sees, not the one the admin sees — and say which surface you checked, because “media library alt fields” and “img tags in the rendered HTML” are different questions with different answers.

How to know it worked

Score identity and discovery separately. Ask the engines “who is X” — that is identity, and correcting sources fixes it. Then ask them the question a buyer would actually type, describing the problem rather than the person — that is discovery, and only placements fix it.

In the August run those two came back as 4 out of 4 and 0 out of 8. A single “AI visibility” verdict would have hidden the entire growth opportunity.

Then re-run the same fixed prompt ladder in thirty days and report the delta. The first run is a baseline, not a result.

Worked example, fully public: the magic berry we assembled for George Leith — 43 verified sources, scored, with where each one goes. The method behind it: the Magic Berry Challenge.

Originally published August 22, 2026.