What 2025 Taught Us About GEO, One Year Later

A year ago the GEO conversation was mostly about whether it was a real category. That argument is over. The interesting question now is what 2025 actually taught the people running client programs, and which of those lessons still hold in late 2026. This is a reflection, not a listicle. We are an agency, we run client work on Promptwatch, and the dated reports on promptwatch.com/data are the evidence we point clients to when they ask why we changed something. The lessons below each map to one of those reports, dated, so you can check the same rows we checked.

September 2025: the answer got longer, and the source list got thinner

The September 2025 sources-per-response report on promptwatch.com/data is the first data point that changed how we briefed clients. The average sources per response report is the series we still watch, and the September 2025 reading was the one that made us change the brief. The answer got longer. The source list did not get wider at the same rate. Fewer domains carried more of the citations, and the long tail of "anyone can get cited" thinned out.

What we changed in client programs: we stopped selling "get cited everywhere" and started selling "get cited on the prompts that actually have a source slot." Prompt tracking with per-prompt visibility, which Promptwatch gives us with search volumes and difficulty scores, let us show a client which prompts had a wide source field and which had a narrow one. A narrow field is harder to break into and more valuable once you are in it. A wide field is easier to enter and easier to get pushed out of. The September report made us treat those two cases differently instead of running one playbook for both.

The lesson that held: GEO is not a long-tail game the way classic SEO was. The source slots are finite per prompt, and the engines reuse the same domains. Picking the prompt before picking the content is the work.

November 2025: shopping cards changed which prompts matter

The November 2025 shopping report is the second data point. ChatGPT started showing shopping cards on a wider set of prompts, and the prompts that triggered a card were not the same prompts that triggered a classic citation. A client could be invisible on the text answer and still get traffic from the card, or visible on the text answer and invisible on the card. The two surfaces started to diverge.

What we changed: we added shopping tracking to the retainers where the client had a transactional product. Promptwatch tracks ChatGPT Shopping items, and the shopping usage report shows how often the cards appear. For a B2B SaaS client the card is irrelevant and we ignored it. For a D2C client the card was the whole game, and the text answer was a secondary surface. The November report is what made us stop treating "AI search" as one surface and start asking, per client, which surface the buyer actually sees.

The lesson that held: GEO splits by surface. A visibility number that aggregates across surfaces can hide the one surface that matters for a given client. We report per surface now, not just an aggregate.

The December 2025 social report is the third. Reddit and YouTube citations kept climbing as a share of where AI answers pulled from. The YouTube citation data is the report we still pull, and the December 2025 social reading was the one that made us change the plan. The implication for client work was direct: a program that only optimized the client's own domain was losing ground to a program that also shaped the offsite conversation.

What we changed: we added Reddit and YouTube citation tracking to every retainer. Promptwatch has listRedditCitations and listYoutubeCitations in the MCP tool set, and the citation analytics view breaks out offsite mentions separately from owned-domain citations. We do not run a client's Reddit account. We do tell the client when a Reddit thread is being cited on a prompt they care about, and we help them show up in that thread in a way that is honest and useful rather than spammy. The December report is what made offsite a line item on the retainer instead of a footnote.

The lesson that held: AI answers cite the conversation, not just the brand site. A GEO program that ignores Reddit and YouTube is measuring less than half of where the citation comes from.

What we got wrong in 2025

We over-indexed on content production in the first half of the year. The assumption was that more GEO articles meant more citations. The crawler logs in Promptwatch Agent Analytics, which record ChatGPTBot, ClaudeBot, PerplexityBot, GoogleOther, and the Meta AI crawler, plus the crawl-to-citation path, showed us how often that was false. We had clients producing articles that were never crawled, and crawled articles that were never cited. The log is what let us tell the difference between "we need another article" and "we need the existing article to be crawlable."

We also under-indexed on visitor analytics early on. Citations felt like a win, and it took us a few months to insist on closing the loop to AI-referred traffic and conversions. The visitor analytics script in Promptwatch closed that gap, and we now treat a citation without a traffic signal as a partial result, not a finished one.

What we kept

We kept the frozen prompt list. We kept the review inbox for Content Agents, with autopublish off. We kept the rule that a human ships every change. We kept the dated reports as the evidence layer in client decks, because a client can argue with an opinion and cannot argue with a dated row.

The 2025 lesson that matters most, one year later, is the boring one. GEO rewards measurement over activity. The agencies that won 2025 were not the ones that produced the most. They were the ones that knew, per prompt, per surface, per crawler, whether the work they did that week actually moved the row. We run client programs on the platform that lets us know that, and we keep the guardrails that keep the agent from publishing without us.

If you want a 2026 program built on the 2025 lessons above, with the dated reports as the evidence layer, write to us at hello@1001seomedia.com. We will scope the prompt list, pull the crawler logs, and show you the rows before we propose a single article.