Measure GEO: citation share tracked with free panels
How to Measure GEO: Track Citation Share With Free Panels

Key Takeaway

  • 📏 Measure GEO like you measure SEO: one number per funnel. Rankings and clicks measure the classic funnel; citation share — how often AI answers name you — measures the answer funnel. Two scoreboards, one routine.
  • 🟦 The free stack is now first-party: Google’s own Generative AI performance report is live in Search Console for every site (Aug 31, 2026), and Bing’s AI Performance report tracks citations in Copilot — both free, both official.
  • 🧪 The protocol: a fixed list of 20–40 buyer questions, polled monthly across ChatGPT, Perplexity, and Gemini, scored for mention rate and citation rate — a spreadsheet replaces a paid dashboard at small scale.
  • 🔁 The loop that compounds: baseline → rewrite the cited-but-beat pages → poll again → record the delta. Measure GEO like an editor, not a spectator: measurement without a follow-up action is expensive history.
  • ⚠️ Know the limits: the Google report counts impressions only (no clicks, no queries), samples skew, and one August 2026 logging bug proves why every number needs a sanity cross-check.

How to measure GEO is the question that keeps the discipline honest. You can feel cited — a friend says the AI mentioned your brand once — or you can measure it: a fixed prompt list, a repeatable routine, first-party panels, and a number that moves month over month. Since August 31, 2026, the measurement problem has first-party plumbing that did not exist last year: Google Search Console now ships a Generative AI performance report to every site, Bing’s Webmaster Tools reports citations in Copilot, and the rest of the answer surface needs nothing more than discipline and a spreadsheet. This guide is the exact measure GEO routine our own low-authority site runs — the free stack, the polling protocol, the math, and the traps.

What GEO Measurement Actually Tracks

GEO measurement answers one question repeatedly: when a user asks something your business should answer, does the AI surface show your brand or your pages? Everything else is decoration around that question. It decomposes into four measurable parts: reach (are you present in the AI answer at all — mentions), sourcing (are your pages cited as sources), accuracy (does the engine describe your brand correctly), and sentiment (is the description favorable). Each has a different method, and mixing them up is how teams end up with numbers that impress nobody who understands them.

The vocabulary matters less than the sampling discipline. “We got cited!” is an anecdote; “we were named in 12 of 30 recorded answers this month, up from 8” is a measurement. Our own routine produced its first real insight on day one: the engines’ answers about our categories consistently cited reference-style guides while skipping opinion pieces — which is a content strategy delivered as a metric, and it cost nothing to discover.

The First-Party Panels: Google and Bing Reports

Google: the Generative AI performance report. Available to every property since August 31, 2026, this report under Performance — the official help page describes it — shows how often links to your site appeared in AI Overviews and AI Mode, breakdowns by page, country, device, and date, with a companion Discover variant. Read its limits like your life depends on it, because the misreadings are everywhere: it counts impressions only — no clicks, no CTR, no position, and no queries — and its impressions are already included in your main Performance report (Google’s AI features documentation confirms the Web-type overlap)’s Web type, so never add the two together. The honest way to read it: as a share and a trend. If your Web impressions total rises while the Generative AI share of it grows, the answer layer is finding you — and the pages tab tells you exactly which pages earn it. A companion setting — Search generative AI control — lets a property opt out of AI features entirely — leave it on Include unless exclusion is a legal strategy.

One 2026 trap deserves its own paragraph: mid-August, AI impressions appeared to collapse across many properties — a logging bug, confirmed by Google, not a real visibility drop, and the reason the routine below says cross-check every unusual number. Trends in a beta panel are directional. Anchor them to your total impressions, your analytics, and your own prompt polls before treating a swing as a verdict.

Bing: the AI Performance report. Bing Webmaster Tools ships a public-preview AI Performance report that tracks citations, cited pages, and grounding queries across Microsoft Copilot and Bing’s AI summaries. Two first-party windows now exist for the two biggest search-based surfaces — and both cost nothing. What neither gives you is ChatGPT, Perplexity, or Gemini: for those, the measurement is yours to run, which is the polling protocol below.

The Polling Protocol: Your Own Citation Survey

Step 1 — build the prompt list. Write 20–40 questions your buyers would ask an assistant: category questions (“best X for Y”), comparison questions (“X vs Z”), trust questions (“is brand X reliable”), and problem questions (“how do I fix X”). Include the queries where you want presence and the queries where competitors currently win — the loss list is the growth area. Our guide to getting cited by ChatGPT and Perplexity lists the question archetypes this list should cover.

Step 2 — poll monthly, same conditions. Ask each list to ChatGPT, Perplexity, and Gemini on a fixed day each month, logged identically: fresh sessions, the same phrasing, results recorded as they appeared. Log per engine: was our brand mentioned (mention), was a page of ours cited as a source (citation), was a competitor named instead (loss), was the description accurate (accuracy note).

Step 3 — compute the two honest rates. Mention rate = answers mentioning your brand ÷ answers tested. Citation rate = answers citing your pages ÷ answers tested. Always report the sample size beside the rate — “40% of 30 tested” is honest; “40%” alone is theater. Track the same two numbers for your top competitor to see relative position, and keep a per-question breakdown so the loss list converts directly into next month’s content queue.

Step 4 — act, then re-measure. Every loss and every wrong description becomes a named action: a page to rewrite answer-first, a claim to publish, a fact to fix, an llms.txt entry to correct. Poll again in 30 days. The GEO for WordPress setup guide covers the technical moves that fix the found gaps, and citation polling takes minutes a month once the list exists.

The Math: Sample Sizes, Rates, and Honest Thresholds

Measurement earns trust through arithmetic, so run the arithmetic early. With a 30-question poll, each answer swings your rate by about 3.3 points; with 40 questions, 2.5. That means a five-point monthly move on a 40-question list is the first number you should take seriously, a ten-point move is a verdict, and anything smaller needs a repeat before it means anything. Per-engine rates compound the issue: ChatGPT, Perplexity, and Gemini each deserve their own column, because their retrieval layers differ and one engine’s answer is not a market signal.

Stratify the list by intent so the rates mean something: 10 category questions, 10 comparison questions, 10 trust questions, 10 problem questions. A 20% overall citation rate that hides “cited on every comparison, absent on every trust question” is a diagnosis, not a score — and the diagnosis writes next month’s content plan. Add a competitor column for the two or three names that dominate your category’s answers, then compute their citation share beside yours: relative citation share is the number that tells you whether the category is winnable at all, and watching two competitors trade the top slot across months teaches the texture of the surface faster than any blog ever will.

Track accuracy separately from presence. An engine that names you but describes the brand wrong is a data-fix task (llms.txt, schema facts, a canonical claim on your about page), not a content task — and conflating the two wastes rewrites on problems that were never about writing. Keep an accuracy column with yes/no/wrong-detail, and treat every “wrong-detail” as a specific filing duty: the fix goes where the model reads brand facts from.

What to Do With a Loss: the Fix Map

Every measured loss maps to one of four fixes, and knowing which one prevents the classic waste of rewriting pages that were never the problem. Loss type one: a competitor cited, your page ranked in the top 20 but not sourced — extraction rewrite territory: answer-first openings, claim-dense tables, FAQ schema. Loss type two: your page sourced but described inaccurately — a facts fix at the entity layer: llms.txt, schema, and a canonical self-description the models can quote. Loss type three: neither you nor a competitor cited — the question resolves from aggregate sources; become the named source with a guide that answers it outright, with numbers. Loss type four: you rank nowhere and are cited nowhere — that is an upstream SEO task, not a GEO task, and the honest answer is to route it back to the classic funnel.

The fix map matters because teams fail the loop at this step: they poll once, see losses, and respond with generic rewrites. A quarter of measured work on the right fix type outranks a year of scattered effort — and the next month’s re-poll grades the choice within 30 days, which is the whole point of any measure GEO routine in the first place.

Reading the Numbers Together

The three sources triangulate: Google’s Generative AI report gives search-embedded AI visibility; Bing’s report gives the Copilot surface; your poll gives the chatbots and, crucially, competitor context. Our own account runs all three, and the numbers tell an aligned story — search AI impressions trending up alongside poll citation-share on reference guides, with the biggest gaps landing on shopping-adjacent queries where our lane is thin. To measure GEO properly you keep all three sources alive. That last finding is exactly why: the panels disagree, and the disagreement is the roadmap.

Set thresholds before you trust movements. With a 30-question sample, one answer is roughly three points — so a move smaller than five points is noise until it repeats. Log monthly, compare quarterly, annotate any platform change (report rollouts, model updates) in your own sheet so future you reads the dips correctly. The evidence-led verdict on SEO in 2026 covers why both funnels deserve their own scoreboards in the first place.

The Paid Options, Honestly

Commercial citation trackers automate the polling across more engines and larger prompt sets — vendor tiers currently run roughly $29 to $499 per month — and our tools ledger reviews what each class actually does. The honest threshold: buy automation when the prompt list outgrows a spreadsheet and the routine has survived three months, not before. First-party panels plus a manual poll answer most small and mid-size needs at zero cost, and they teach you what the dashboards are actually claiming — which is the knowledge that makes you a good purchaser, if you ever become one.

Frequently Asked Questions

How do I measure GEO?

Three free sources: Google Search Console’s Generative AI performance report (impressions in AI Overviews and AI Mode, by page and country), Bing Webmaster Tools’ AI Performance report (citations in Copilot), and your own monthly poll of 20–40 buyer questions across ChatGPT, Perplexity, and Gemini, scored as mention rate and citation rate with the sample size stated.

What is citation share in GEO?

Citation share is the share of relevant AI answers that source your material — citations per answers tested, computed per engine and per question set. It is the GEO scoreboards’ equivalent of click-through rate: not the only number, but the one that says the answer engine is choosing you.

Does Google Search Console show AI Overview traffic?

The Generative AI performance report shows how often your links appeared in AI Overviews and AI Mode — impressions only, grouped by page, country, device, and date, with those impressions already included in the main Performance report’s Web type. Clicks from those features are recorded in Performance but queries are not in the AI report.

How often should I run a GEO citation poll?

Monthly, on a fixed day, with the identical question list. Weekly polling wastes effort — the surfaces move slower than that — while quarterly loses the cause-and-effect link between your changes and the next result. Thirty minutes per month is the honest cadence.

What is a good citation share?

There is no official benchmark, which is why the sample size matters more than the percentage. From our own tracking: any non-zero citation share on your own category questions is the first milestone; a double-digit share on comparison questions means the surface knows your brand; and for a small site the honest measure GEO target is simple — beat last month, then beat your strongest competitor. Absolute numbers move as engines update; relative position is the durable metric.

Can I measure GEO for ChatGPT and Perplexity without paid tools?

Yes — that is exactly what the polling protocol does: a fixed list, fresh sessions, identical phrasing, logged results, and two rates computed monthly. It is manual, it is free, and at a 20–40 question sample it is as truthful as most dashboards, because you see the raw answers yourself.

The Verdict

How to measure GEO in 2026 has an honest answer at last: two first-party panels for the search-embedded surfaces, one disciplined monthly poll for the chatbots, two rates computed with the sample size stated, and a threshold before you call any move real. The free stack measures everything a small or mid-size site actually needs — and the number that matters is not a dashboard score. It is whether the AI answer names you, and whether that share grows after you do the work.