Campaigns · 8 min read

Brand GEO campaigns: the publisher's A-to-Z playbook

Anton Nashkerskyi · ToldBy · August 12, 2026

A brand GEO campaignis a measured effort to change how AI assistants answer the questions that matter to a brand. You fix a short list of those questions, check how often each engine’s answers mention the brand today, publish content aimed at the gaps, and re-measure on a schedule until the numbers move — or until you can say, with evidence, that they didn’t. And the party best equipped to run one isn’t the brand. It’s the publisher the brand already buys media from: the trusted, constantly-crawled editorial pages AI answers draw on are the pages publishers own.

Why publishers, and why now

People increasingly ask instead of searching. In OpenAI’s own usage study with NBER researchers, 49% of ChatGPT messages are “Asking” — seeking information or advice — versus 40% asking the model to do a task (Chatterji et al., NBER w34255, 2025). When an assistant answers one of those questions about a category your advertiser competes in, it typically names a handful of sources and reads them out as fact. Either the brand is in that answer, or a competitor is.

Brands can’t easily fix this alone. What they can publish about themselves is marketing; what moves an answer about “which sunscreen won’t break me out” is the kind of tested, editorial, regularly-updated coverage that publishers produce every day — in verticals where the publisher already has years of topical authority. That makes a brand GEO campaign a natural publisher product: the advertiser brings the subject, the publisher brings the pages AI already trusts, and the campaign is sold, run and reported like any other media product — with better measurement than most. (The usual rules of commercial content apply: disclose sponsorship the way you would in print. Nothing about AI changes that.)

Track questions, not prompts

The first design decision is the one most tools get wrong. If you track prompt strings (“best mineral sunscreen 2026 site:reddit”-style artifacts), every rewording breaks your history, and after a month you’re comparing apples to whatever the intern typed. A campaign should own a short, fixed list of questions — the real things users want to know about the subject:

• Branded questions — the user already knows the subject. “Is Aurelle safe for sensitive skin?” Rolled up, these measure awareness: when people ask about you, does AI know you?
• Category questions — the user is asking around the topic without naming anyone. “What sunscreen won’t break me out?” Rolled up, these measure surfacing: does AI bring you up unprompted? This is where campaigns are won.

Each question gets a weight (how much it matters commercially) and a handful of disposable wordings— the actual strings sent to engines. Wordings can be regenerated, edited or swapped at any time without breaking trends, because the tracked unit is the question. Lock the question list and you’ve built the campaign’s measurement instrument; every later run is comparable to the baseline.

Ask each engine the way its users ask

One question, phrased two ways. Semrush’s analysis of 80 million clickstream records found ChatGPT prompts average ~23 words in normal chat but drop to ~4.2 words when a search surface is involved (Semrush, 2025). So a campaign probes Google’s AI Overviews and AI Mode with short, search-style wordings, and ChatGPT and Copilot with conversational ones. Engines are then compared on the same questions, each phrased the way its users actually phrase them — field realism, not identical-string lab conditions.

The playbook, A to Z

AAnchor the subject

Name the thing the campaign is about — a brand, a product line, an event — plus its aliases (short names, abbreviations, common misspellings). The subject, not any single URL, is what every number in the campaign refers to.

BBuild the knowledge base

Collect what the public web currently says about the subject: facts, dates, claims, and which sources carry them. This takes minutes with web search, but verify the facts with the client before they anchor the campaign — if the knowledge base has the wrong launch date, your questions will inherit it.

CCurate and lock the questions

Generate candidates from the knowledge base, then edit like an editor: ~12 questions, roughly 40% branded / 60% category, every commercially-relevant subtopic covered, each weighted 1–3. This half hour of curation is the highest-leverage work in the whole campaign. Then lock the list.

DBaseline every engine

Probe each engine with each question’s wordings, several times each — one answer is an anecdote; repetition is a measurement. Score mentions deterministically (does the answer name the subject or an alias — no AI judging AI), and keep every raw answer so any number can be audited later.

ERead the table

The campaign’s centerpiece is one table: questions down the side, engines across the top, each cell the share of answers that mentioned the subject. Here’s an illustrative baseline for a fictional sunscreen brand, Aurelle, whose campaign a beauty publisher is running:

Illustrative example · fictional brand · cells show mention rate + sample size

Read it like heat, not verdicts — cells are small samples; confidence intervals belong on the rollups. The story above: awareness is respectable (AI mostly knows Aurelle when asked directly), but surfacing is near zero— on the high-weight category questions, the brand doesn’t exist. And for every weak cell you also see the gap fillers: which domains ARE being cited there. That short, ranked list — high-weight, low-rate cells plus who currently owns them — is the campaign brief.

FMake the editorial moves

Now the publisher does what only a publisher can: commission and update the coverage that answers the weak questions. In the Aurelle example, the dermatology vertical publishes its tested round-up of sunscreens for reactive skin — the page that actually answers “what sunscreen won’t break me out?”— and refreshes the existing comparison pages where the brand legitimately belongs. Keep the pages crawlable to search and AI bots, and skip the tricks: in our own testing and the published evidence, hack-style tactics (llms.txt files, schema stunts, keyword stuffing) don’t survive scrutiny. What survives is the boring pair — content quality and crawler access.

GProve it (or report the null)

Re-probe on a schedule against the locked question set. Because the instrument is frozen and every answer is stored, the trend line is honest: same questions, same scoring, engine by engine, with the question-set version stamped on the chart. That is the artifact you put in front of the advertiser — we don’t promise outcomes, and neither should you. AI engines change under everyone’s feet. What a campaign guarantees is that you’ll know what moved, what didn’t, and what to try next — with numbers a skeptical client can audit down to the raw answers.

Common questions

Isn’t this just rank tracking for AI?

No. Rank trackers watch one URL’s position for keyword strings. A campaign watches a subjectacross a locked set of user questions, in each engine’s native register, scored by mentions in the answer text itself — because in AI search, being in the answer matters more than being in a list of links.

Can you guarantee the mention rate goes up?

No — and be suspicious of anyone who does. Engines re-rank, models update, and some gaps are structural. What’s guaranteed is honest measurement: a locked instrument, deterministic scoring, stored evidence, and a trend that includes the results that didn’t go your way.

When can I run one?

Campaigns are rolling out in early access with design partners. Request a demo and mention Campaigns, or start with a free scorecard on your own pages to see the measurement pipeline first. The product overview lives at toldby.io/campaigns.