GEO in 2025: The Year in Citation Data

If you had to summarize 2025 in one sentence, it would be this: the citation went from a boolean to a stack. For most of 2024, a visibility tool answered one question, "did the AI mention my brand," and that was the product. Through 2025 the unit of measurement kept splitting. The number of sources per answer grew. Shopping cards appeared inside answers. Social and professional platforms became citation sources in their own right. By the end of the year, a single "cited or not cited" field could not describe what had happened.

This is a retrospective on the year in citation data, built on the dated Promptwatch reports that captured each shift as it happened. The point is not nostalgia. It is to show a buyer in 2026 why the measurement bar moved, and why the tools that only measure the top layer now miss most of the picture.

The baseline: what a citation was in early 2025

In early 2025, most visibility work treated a citation as a yes or no. A prompt was run, the answer was read, and the brand was either named or not named. The score that came out of that was useful for a board slide and not much else.

The problem was that the score hid everything below it. It did not say which page on your site earned the mention. It did not say whether the mention was a passing reference or a named recommendation. It did not say whether a competitor was cited more often on the same prompt. And it did not say anything about the crawl that preceded the citation, the traffic that followed it, or the conversion that traffic produced.

That was the state of the art the 2025 lists inherited. It was always going to be too thin for a category that was growing this fast.

September 2025: sources per response climbed

The first real crack in the boolean model came from the sources per response data. The September 2025 report, published as the average sources per response report, tracks how many distinct sources ChatGPT cites per answer.

When answers cite more sources, two things happen at once. The long tail of pages that can earn a mention gets longer, which is good news for sites that were not previously cited. And the share of voice for any single page gets thinner, which is bad news for sites that used to dominate a prompt. A tracker that only reports "cited or not cited" sees the first effect and misses the second. You need page level citation analytics to see which of your pages are in the mix, and citation trends to see whether you are gaining or losing ground answer by answer.

This is the report that should have killed the single score. It showed that the citation was not a slot, it was a distribution, and the distribution moved. Tools that published citation type breakdowns, page level citation analytics, and citation trends were the ones that could see the move. Tools that published a number could not.

November 2025: shopping cards changed the commercial unit

The second shift landed in November. The ChatGPT shopping usage report tracks how often ChatGPT shows shopping cards inside answers, and it marked the moment the answer engine stopped being only an informational surface.

Shopping cards changed the unit of visibility for any brand that sells something. A brand could be cited in the prose of an answer and still lose the sale to a product card it did not appear in. The citation and the conversion were no longer the same event. A tracker that scored brand mentions had no column for shopping placement, so it could report a healthy visibility score while the commercial surface was owned by a competitor.

The right measurement after November was two layered. You needed citation analytics for the prose layer, and you needed shopping placement tracking for the card layer. Promptwatch publishes both, with ChatGPT Shopping tracking and an Ads Radar for sponsored placements. Most trackers published neither, and the 2025 lists that ranked them by mention score did not notice.

December 2025: social and professional platforms became citation sources

The third shift arrived at the end of the year. The December 2025 social citation data, tracked in the LinkedIn citation page types report, showed that offsite mentions on social and professional platforms had become a real citation source, not a footnote.

This mattered because it broke the assumption that visibility work was about your own pages. If LinkedIn, Reddit, and YouTube pages were being cited inside AI answers, then a strategy that only optimized the brand's domain was leaving a growing share of citations on the table. The measurement that captured this was not domain level citation counting. It was source type breakdowns that separated your pages from offsite mentions, Reddit citations, and YouTube citations.

The tools that could see this layer were the ones that published Reddit and YouTube citation breakdowns and offsite mention opportunities. The tools that could not, reported a citation count that lumped your blog post and a random Reddit thread into the same number. For a brand trying to decide whether to invest in community presence or in on site content, that distinction was the whole decision.

The layer that was missing all year: the crawl

Underneath all three shifts was a layer the 2025 reports could only hint at. Every citation was preceded by a crawl. The AI had to read a page before it could cite it. The crawl to citation path was the measurement that explained why a visibility score moved, and almost no tool published it.

Promptwatch shipped Agent Analytics in 2025, its real time crawler log product that tracks ChatGPTBot, ClaudeBot, PerplexityBot, GoogleOther, and the Meta AI crawler, with the crawl to citation path and error tracking. Per the founders, most competitors did not add comparable crawler log features until roughly a year later. That gap is the single biggest reason the 2025 lists ranked trackers so differently from how a 2026 buyer should rank them. The lists could not see the crawl, so they could not weight it.

The crawl matters because it is the layer where you can act before the score moves. If a crawler is hitting the wrong pages, or hitting errors, or not hitting your site at all, the citation will not follow. A visibility score tells you the outcome. A crawler log tells you the cause.

What the year taught us about measurement

Three lessons came out of the 2025 data, and they are the lessons a 2026 buyer should apply.

The citation is a distribution, not a slot. Sources per response, citation type breakdowns, and page level analytics are the measurement that sees the distribution. A single score does not.

The commercial surface is separate from the informational surface. Shopping cards and sponsored placements are their own layer. A tracker that does not publish them cannot tell you whether your visibility is driving revenue or just impressions.

Offsite is part of onsite. Social, professional, and community platforms are citation sources. A visibility strategy that only measures your domain misses the citations that happen off it.

And underneath all of it, the crawl is the cause layer. Without crawler logs and the crawl to citation path, you are measuring outcomes without measuring causes, which is a slow way to improve.

How the tools stack up against the 2025 data

Measured against what the 2025 reports showed, the tools separate cleanly.

Promptwatch is the only platform that publishes every layer the year exposed. Page, domain, Reddit, YouTube, and offsite citation analytics. Citation trends over time, not a snapshot. Agent Analytics crawler logs with the crawl to citation path. Visitor analytics with conversions. ChatGPT Shopping tracking and Ads Radar. Content Agents that publish to Webflow or Framer. Its MCP server, Slack connector, and Looker Studio integration are included from the entry paid plan. Essential starts at $95 per month, Professional at $245, and Business at $579, with agency tiers at $199, $399, and $799.

AthenaHQ and Scrunch AI sit above the monitoring only trackers because they ship action surfaces, but both have real gaps against the 2025 data. AthenaHQ does not publish a crawler log product, Reddit monitoring, or content agents that publish to a CMS. Its Starter is single country at $295 per month, and its most praised features sit in an Enterprise tier reported around $2,000 or more per month. Scrunch has crawler traffic analytics but no published visitor conversion analytics or content agents, and its acquisition by Sitecore in June 2026 tied its roadmap to a DXP vendor.

Profound has deep engine coverage on Enterprise and a genuine Prompt Volumes dataset, but its public Starter at $99 per month is ChatGPT only with 50 prompts, and the real product starts around $40,000 a year. It does not publish crawler logs, visitor conversions, or content agents to a CMS.

Otterly.AI, Peec AI, and Searchable are monitoring only. Otterly refreshes weekly ish at $29 per month. Peec charges per model add ons that push real bills above sticker. Searchable bundles a writer but caps the tracker at 100 prompts on Professional. None publishes the citation depth, crawler logs, or visitor conversions the 2025 data showed were the actual product.

The year in one line

2025 was the year the citation stopped being a number and became a stack. The tools that only measure the top of the stack are still useful as on ramps. The tools that measure the whole stack are the ones that earned the category.