AG✦

Entrepreneurship

Old Press Does Not Expire, It Compounds: Auditing What AI Still Reads About You

By admin
November 19, 2026

Field Notes

Language models read your whole published record at once, so a 2019 bio with a stale title is live testimony about who you are.

Search ages content, retrieval does not

A page that stops ranking stops getting traffic. It does not stop existing in the corpus a model trained on or the index a model retrieves from.

Your old press is still answering questions about you. You have stopped reading it, which is a separate matter from whether it is gone.

Ranking is a weak predictor of what gets cited

Semrush‘s AI search traffic study, published July 21, 2025 across more than 500 digital marketing and SEO topics, found that pages cited by ChatGPT search rank in traditional positions 21 or lower almost 90 percent of the time. The study does not disclose its prompt count or its visit counts.

A 2019 interview sitting on page three of Google is well inside that band. Position tells you little about whether a model will quote the page.

The same study reported that about 50 percent of links in ChatGPT 4o responses point to business or service websites, and that an average visit from AI search is 4.4 times as valuable as an average organic visit by conversion rate. Semrush did not publish the underlying conversion rates, so treat that multiple as directional.

Where the citations concentrate

Profound’s “AI Platform Citation Patterns,” published June 5, 2025 and updated in August 2025, analyzed 680 million citations from August 2024 through June 2025. Profound does not disclose its query sampling method, which limits how far you should generalize from it.

Within ChatGPT’s top 10 sources, Wikipedia accounted for 47.9 percent, Reddit 11.3 percent, Forbes 6.8 percent, G2 6.7 percent and TechRadar 5.5 percent. Within Google AI Overviews’ top 10, Reddit took 21.0 percent, YouTube 18.8 percent, Quora 14.3 percent, LinkedIn 13.0 percent and Gartner 7.1 percent.

Profound also split ChatGPT citations by top level domain: .com at 80.41 percent, .org at 11.29 percent, .uk at 2.16 percent, .io at 1.67 percent and .ai at 1.13 percent.

LLM Pulse‘s top cited domains list, updated October 26, 2026 on a seven-day rolling window, puts youtube.com at 9.73 percent, google.com at 6.25 percent, reddit.com at 5.97 percent, facebook.com at 4.92 percent, instagram.com at 3.33 percent, linkedin.com at 2.86 percent and trustpilot.com at 2.09 percent. The page carries an unresolved caveat, since its chart heading says last 90 days while the methodology says seven days, and the sample size is not published.

The corroboration loop runs in both directions

Jason Barnard, writing in Search Engine Journal on June 28, 2021 and modified October 17, 2021, described the mechanism as an Entity Home, a single authoritative page you own, marked up with schema.org, then corroborated by authoritative first and third party sources that link back to it. He called the result an “infinite, self-confirming loop,” and noted that the Knowledge Graph has no notability guidelines.

A loop that confirms correct facts confirms stale ones by the same route. An old bio listing a company you left in 2021 is corroboration, and it is pulling against your current description of yourself.

What I audit, field by field

I run Subscribe PR in Phoenix, and the first pass on a new engagement is an inventory rather than a strategy document.

I pull the pages that mention the person by name, then score each one on five fields: name spelling, current title, company name, city, and any credential or affiliation. A page that disagrees with the Entity Home on two or more fields goes on a contradiction list.

That list gets sorted by domain type, because a stale fact on Wikipedia or a major publisher carries more weight in the citation data above than the same fact on a dormant blog.

Where the contradictions come from

Former job titles are the most common source, followed by company names that changed after an acquisition.

Conference bios are the quiet problem. A speaker page from 2018 often outlives the company it describes, keeps its URL, and gets no correction request because the person forgot it existed.

Bylines under a maiden name or an earlier professional name split one person into two entities, and a model has no reliable way to merge them without a page that states the connection in plain text.

Correct before you try to delete

Removal is the slowest lever and the least available one. Correction moves faster, because a site owner updating a job title has no editorial objection to overcome.

Where you control the page, update it and date the update. Where someone else controls it, send the specific field and the specific replacement instead of a general complaint.

Then add new corroboration, so the current version of each fact appears on more pages than the stale version does.

Publishing new material on top of an uncorrected record adds volume to a contradiction. Run the inventory first and you know which pages the new work has to outweigh.

If you would rather see how that inventory gets run, Caleb Hellinger sets out the method in the write-ups at calebhellinger.com/insights.

Caleb Hellinger runs Subscribe PR, an answer engine optimization and digital PR agency in Phoenix.