Digital PR and GEO: why AI brand mentions belong on the brand scorecard

Is GEO just SEO version two?

No. The medium differs, the number that proves success differs, and the buyer's journey through it differs, and those three differences move GEO out of the performance team's frame. SEO competes for a position in a list of links a person then clicks. GEO competes to be named inside a written answer that person may never click away from.

The click data is the plainest way to see it. In the first four months of 2026, 68.01 percent of US Google searches ended without a click, up from 60.45 percent in 2024 (SparkToro and Similarweb clickstream analysis, Rand Fishkin). The same study puts AI Overviews on 20 percent or more of searches and finds they cut click-through rate by close to 60 percent where they appear. Click-through rate is the share of people who see a result and click it. What still reliably sends a click is a person typing your name: branded terms are roughly 44 percent of Google searches (SparkToro, Amanda Natividad).

Read those together and what the channel rewards changes. If two thirds of searches resolve on the page, being described well inside that answer is the win, and the click comes later, under your own name. Tim Soulo, Ahrefs' chief marketing officer, put it in one line: the game is shifting from counting clicks to tracking how often Google's AI mentions your brand.

This is also why the reporting feels broken when a performance team runs it. Google Analytics 4, or GA4, has no channel called AI search. Demand created inside an assistant lands as direct traffic, branded search and homepage visits, so a last-click dashboard shows a flat line while the work is happening. That framing comes from practitioner Matthew Mellinger with no dataset attached, so read it as a description of the tooling. Our guide to tracking AI referral traffic in GA4 covers what the tool can and cannot see.

One vocabulary note. AEO, answer engine optimisation, is used almost interchangeably with GEO, and our AEO vs GEO vs SEO explainer sets out the split, while our guide to generative engine optimisation covers the mechanics. Which acronym your agency prefers changes nothing here. Who in your building is accountable changes everything.

What does an AI answer actually say about your brand?

Whatever the sources it read say. An assistant asked about your category assembles a description out of third-party material, and part of what it produces is characterisation the brand never wrote: how established you are, who you suit, what you cost too much for. If you have not put your own account of yourself into the places the model reads, it builds one from whoever did.

That is the brand-side problem in one sentence, and it is a communications problem before it is a technical one. On the same stage, Christian Ward of Yext presented two studies covering 38 million data points showing that memory and personalisation now rewrite what a given person sees, so the same brand renders differently from one context to the next. A screenshot of one assistant on one day understates that variance badly, which is why a tracked trend is the usable artifact.

The practical consequence is narrative control, and it is ordinary brand work. Describe the company in the exact words you want an assistant to repeat, and publish those words where third parties will pick them up: the boilerplate, the founder interview, the analyst briefing, the category definition in a bylined piece. The phrasing that travels is the phrasing that is easiest to lift: a journalist reuses it, and the model reads the journalist. A brand that has agreed one description and repeated it across earned coverage feeds a consistent input into every answer written about it. Where the descriptions conflict, the model reaches for whichever competitor stated it once, clearly.

Where do AI citations actually come from?

Earned media, by a distance, and the brand's own site is a small minority of the total. The largest public dataset puts earned media at 84 percent of all AI citations, journalism alone at 27 percent, and paid or advertorial content at 0.3 percent, across more than 25 million citations from ChatGPT, Claude and Gemini in 17 industries (Muck Rack, Generative Pulse, May 2026). Across three editions since July 2025 the earned share has sat between 82 and 89 percent, so the finding has held across three readings of the dataset.

The closest thing to a controlled test comes from Stacker, which took the same articles and changed only where they were published. On the brand's own domain those articles were cited in 7.6 percent of prompt and platform combinations; distributed through third-party outlets, with no edit to the content, they reached 34 percent, across 8 articles, 5 large language models and 944 combinations (Stacker, Noah Greenberg, 9 December 2025). The publisher sells distribution, so treat the size of the gap as indicative and the direction as well supported.

Do brand mentions matter more than backlinks?

Yes, and the gap is large enough to change where a budget goes. Ahrefs studied 75,000 brands and found web mentions correlate 0.664 with AI Overview visibility while backlinks correlate 0.218 (Louise Linehan with Xibeijia Guan, 26 May 2025). The quartile breakdown in the same study is the part worth reading twice: brands in the top 25 percent for web mentions average 169 AI Overview mentions, the next quartile down averages 14, and the bottom half averages 0 to 3. That is not a gentle gradient. Being talked about often is close to a precondition for being named at all.

The reason unlinked mentions count is mechanical. An assistant builds an answer by retrieving passages, resolving the names in them to entities¹ it already knows, and grounding² its sentences in that material. A brand named in a paragraph of an industry roundup feeds that process whether or not the paragraph carries a hyperlink. Measurement practice has caught up: Aleyda Solis's prompt library records linked citations and unlinked mentions as separate fields per tracked prompt, and separates a recommendation from a bare naming, because the three are worth different amounts.

For a brand team this is familiar work under a new name. Coverage volume, message consistency and share of voice are what you already fight for. They now also decide whether an assistant can say anything about you when a buyer asks who the serious options are.

What should a brand team measure for GEO?

Share of answers³ is where the benchmark question lands first. Siege Media publishes bands of above 70 percent as excellent, 40 to 70 percent as good and below 40 percent as needing work (Siege Media, GEO ROI). Siege set those bands and sells against them, so use them to structure the conversation and build your own baseline per category before anyone commits to a target.

Sentiment belongs on the scorecard, and the honest caveat belongs next to it. How an assistant characterises you, warm or lukewarm, current or dated, premium or cheap, is what a brand team is paid to influence, and it moves before anything else does. The measurement method is still young: every vendor codes answer sentiment differently, the public evidence is thin, and the strongest published example is a single unnamed brand in a vendor case study that called its own result directional. Track it, chart the trend, and give it to your board as a read.

Branded search and pipeline are correlation, and saying so protects you. The Instacart case pairs a 461,000 increase in branded search with 46,500 LLM citations in the same period: a strong signal, a weak causal claim. The same source finds 30 to 60 percent of sales self-report discovering a vendor through an assistant, far above anything the analytics show, because assistants strip the referrer. Our guide to measuring AEO covers what each tool can see.

Which of your existing levers move AI visibility, and who owns each?

Most of them, and most are already funded outside SEO. Aleyda Solis makes the point structurally in her framework for prioritising third-party citation work, which scores an activity on six dimensions, and one of those is marketing alignment. Her definition, verbatim: "Marketing alignment assesses how the activity supports and can be aligned with an existing marketing priority or workflow, such as digital PR, community management, creator marketing, reviews and reputation, marketplace management, partnerships or customer marketing" (Aleyda Solis, 20 August 2026). The framework names 13 activity types to pursue and 11 to reduce. The worksheet scores are her editorial judgement, called directional starting points, so take the rubric and score your own cells.

Read that list of workflows again. Digital PR, community, creator, reviews, partnerships, customer marketing. In most marketing organisations six of the seven report to a brand or marketing lead. Naming the owner of each lever is what turns a GEO report into work someone does on Monday.

The technical row goes to the SEO or web team, and it is genuinely their job: pages structured so a single passage answers a single question and stands alone out of context, schema in place, nothing blocking the crawlers that feed the assistants. Everything above it is brand work with a new reporting line. Our GEO service page sets out how the two halves run together.

See how assistants describe your brand today. Book a free AEO audit.

Which digital PR tactics are GEO waste?

The ones you buy for reach. Paid and advertorial content accounts for 0.3 percent of all AI citations in the Muck Rack dataset, against 84 percent for earned media. Syndicated release distribution, paid inclusion in a roundup and a logo on a partner wall each produce a URL and a screenshot for the report, and close to nothing an assistant will quote. They are also the cheapest things to sell, which is why they lead low-end GEO packages.

The same Solis framework lists what to reduce: generic brand mentions and logo walls, reach-first coverage on irrelevant generalist sites, syndicated press releases, thin "best of" listicles and paid inclusions with no stated methodology, promotional guest posts, low-quality directory creation, manufactured reviews, stale profiles, and bulk AI-written articles carrying no original evidence.

The selection rule that replaces reach is audience fit, and there is now a number on it. Fractl, using SparkToro's audience data across eight industries, found niche publishers delivered 1.7 times the audience affinity of major outlets while attracting roughly 130 times less traffic. In their worked example, RecruitingDaily, with about 10,000 monthly visitors and an affinity score of 93, beat GetApp, with 418,000 visitors and a score of 90 (Kelsey Libert, Fractl, on SparkToro, 19 June 2026). One of those placements puts your name in front of the sources an assistant treats as expert in your category.

The press release itself survives this. A genuinely newsworthy announcement that a journalist reports on lands as earned coverage and counts. The distribution wire sold alongside it does not.

Five things a head of brand can do this quarter

1. Put AI visibility on the brand scorecard, in the existing rows. Mention rate under unaided awareness, share of answers under share of voice, answer sentiment under sentiment, branded-search lift under branded search. One new column, no new dashboard.
2. Baseline before you brief anyone. Twenty buyer prompts, three engines, coded for named, recommended or absent, plus sentiment and description accuracy. You cannot brief a PR team against a target you have never measured.
3. Agree one description of the company and get it into earned coverage. The words a journalist finds easiest to lift are the words an assistant repeats.
4. Move budget from reach placements to fit placements. Score outlets on whether their readers are your buyers, and drop syndicated distribution and paid roundup inclusion.
5. Name an owner for every lever. Communications, community, customer marketing, partnerships and insights each hold one. Give SEO the extractability row and stop routing the programme through them.

Find out what AI says about you. Book a free AEO audit.

Frequently asked questions