Summarize this article with AI
Ask with ChatGPTOpen with this article
Ask with ClaudeOpen with this article
Ask with PerplexityOpen with this articleLast Monday, a Head of Demand at a $40M ARR analytics SaaS opened her Profound dashboard, screenshotted a number, and forwarded it to her CEO with the subject line “we exist”. 38% citation shareacross ChatGPT, Perplexity and Gemini. Forty seconds later he wrote back two words: “now what”.
That gap, between a number on a dashboard and a decision a CMO can make on Monday morning, is where most GEO measurement falls apart in 2026. The tools have multiplied. The metrics have not. Marketing teams are paying $99 to $499 a month for a vanity number that nobody can act on, when what they need is a measurement stack that ties citation share to the page that earned it, the engine that surfaced it, and the rewrite that would lift it.
Why citation share is the only metric that survives a model update
Every other GEO metric is a downstream proxy. Traffic dies when ChatGPT changes its retrieval recipe. Impressions die when a model ships a fresher index. Brand mentions die when a competitor writes a better passage. Citation share, the percentage of in-category prompts where your brand is named, is the only number that maps directly to the LLM’s decision.
Discovered Labs calls the alternative the dashboard trap: tools that show you a number every week without telling you which page caused it, which competitor displaced you, or which sentence to rewrite. Backlinko’s 2026 round-up of five visibility tools makes the same point in a different shape. Most platforms are diagnostic. Almost none are operative.
Three reasons citation share holds up where the others fail:
- It is engine-agnostic. The formula reads the same way in ChatGPT, Claude, Perplexity, Gemini, Copilot and Google AI Overviews. A tool that measures it on three engines and a tool that measures it on six are doing the same math, just on a wider surface.
- It is sample-stable. When pulled weekly against a fixed prompt set of 200+ items per category, the noise floor drops below 4 points week-on-week. Below that sample size, the answer churn we describe below swallows your signal.
- It correlates with pipeline. In Profound’s public 2025 cohort data, B2B SaaS customers who lifted citation share by 10 points saw pipeline lifts in the 20 to 40% range one quarter later. We replicated the directional pattern in our own customer set.
A note on what citation share is not
- It is not brand mention count. A “Notion” showing up inside a paragraph about productivity is a mention. A “Notion” that answers the user’s actual question is a citation. The first is noise. The second moves pipeline.
- It is not search volume. The same 100 prompts can return 100 different answer surfaces depending on the model, the index date, and the temperature.
- It is not impressions. ChatGPT does not report them. Pretending to track them is theatre.
The 4 metrics that beat traffic, and the formulas behind each
Most GEO articles in the SERP today (we audited the top 15 last week) name “share of voice” as a metric and stop there. None of them ships a formula. Here are the four we use, with the exact math, the sampling rule, and the cadence.
1. Citation share
Formula: citation_share = cited_prompts / answered_prompts measured over a fixed prompt set.
Sampling rule: 200 prompts per category, refreshed weekly. Below that, the 31% answer-churn rate eats your signal. Above 500, you get diminishing returns and a slower feedback loop.
Cadence: read it weekly, not daily. Models cache results for hours. A daily pull mostly measures cache state.
2. Engine spread
Formula: engine_spread = engines_citing / engines_tracked.
A 70% citation share concentrated on ChatGPT alone is structurally weaker than a 35% citation share spread evenly across six engines. The first is a single-point-of-failure exposure. The second is a durable position. Profound covers eight engines. Otterly covers four. Same number on the dashboard, very different exposure profile.
3. Passage prominence
Formula: a weighted score where citation in the first sentence of an answer counts 3, mid-answer 2, list item 1, no citation 0. Average across the prompt set.
Citations buried in a list at the bottom of a Perplexity answer convert at a fraction of citations that open the answer. Discovered Labs flags this as the difference between being named and being endorsed. The same brand can show identical citation share across two engines and ship 4x the demos from the engine that quotes it first.
4. Answer churn
Formula: churn = 1 - (cited_at_t1 ∩ cited_at_t2) / cited_at_t1 across two runs of the same prompt 48 hours apart.
An answer that includes you 100% of the time is structurally different from an answer that includes you 50% of the time. Both can read as “38% citation share” on a weekly average. Only the first one is real. In our Q1 2026 dataset, the median B2B SaaS prompt churned 31% in 48 hours. If your tool is not measuring churn, you are buying a snapshot of a moving target.
A citation share number you cannot trace back to a passage is not data. It is an alibi.
Run this 5-minute diagnostic before you buy any tool
Before you put $499 a month on a Profound or a Peec, run this diagnostic by hand. It costs zero and tells you whether you have a tool problem or a content problem. Most teams find out it is the second.
Category-defining prompt
Run "What are the best [your category] tools in 2026? Include pricing." in ChatGPT, Claude, Perplexity, Gemini, Copilot and Google AI Overviews. Log who is named first, second, third. This is your baseline citation share at the most commercial query in your space.
Comparison prompt
Run "Compare [your brand] vs [top 2 competitors] for [primary use case]. Cite sources." Log which URLs the engines cite. If they cite your competitors and not you, the gap is content-side, not tool-side.
Buyer-intent prompt
Run "I am a VP Marketing at a $20M ARR B2B SaaS, which [category] tool should I buy and why?" This is the conversion query. The brand named in the answer wins the demo more than 60% of the time in our sample.
Long-tail problem prompt
Run "How do I measure citation share across ChatGPT, Claude and Perplexity?" This tests whether your educational content is passage-shaped enough to be retrieved. Most B2B SaaS sites lose this query to Backlinko, Semrush and the vendor blogs.
Stability prompt
Run prompt 1 three times in a fresh session over 48 hours. Log how often you appear. If your inclusion rate is below 67%, churn is your bottleneck, not citation share.
The 8-tool measurement stack, scored honestly across 6 engines
The GEO tooling market in 2026 is split into two camps: monitoring-first platforms that surface a citation number, and operative platforms that pair the number with a content recommendation. Both have a place. Below is how the eight tools we test against stack up, scored on the metrics above, with verified entry pricing.
| Tool | Engines tracked | Content rec. | Starting price | Best for |
|---|---|---|---|---|
| Profound | 8 (incl. Grok, DeepSeek) | No | from $99/mo | Mid-market with engineering bandwidth |
| Peec AI | 4 to 10 (gated) | Partial | from €89/mo | EU teams scaling 1 to 5 engines |
| Otterly | 4 (CGPT, PPLX, Copilot, AIO) | No | from $29/mo | Solo marketers running a first audit |
| AthenaHQ | Multi-engine | Yes | Custom | Technical SEO specialists |
| Scrunch | Multi-engine | Yes | Custom | In-house content teams |
| Semrush AI Toolkit | Multi-engine | Yes | $99/mo per domain | Teams already on the Semrush suite |
| Superlines | Multi-engine | Yes | Custom | Multi-brand portfolios |
| ClaironUs | 6 (CGPT, Claude, PPLX, Gemini, Copilot, AIO) | Yes | from $49/mo | Teams that want measurement and passage-level rewrites in one loop |
Pricing verified April 2026 against Backlinko’s round-up and each vendor’s public page. “Custom” means no public list price at time of writing.
How to pick (3 quick heuristics)
- Solo marketer running a first audit: start with Otterly at $29 a month. Four engines is enough to baseline.
- Mid-market with a content team: pair Profound or Peec for measurement with a tool that ships passage-level rewrites (Scrunch, AthenaHQ, or Clairon). Measurement-only platforms leave money on the table.
- SEO team already on Semrush: add the Semrush AI Toolkit at $99/mo per domain to your existing seat before adding a separate tool. The integration with rank data is worth the trade-off in engine coverage.
The cadence: weekly, monthly, quarterly
GEO measurement breaks at two ends. Teams either measure too often (daily, which mostly captures cache state) or too rarely (quarterly, which misses the rewrite feedback loop). Here is the cadence we run on our own content and recommend to every customer team.
Weekly: re-run the 200 prompts, log the deltas
The week-on-week delta is the only number worth a Slack notification. A 3-point gain in week 2 is realistic. A 15-point gain probably means your sampling broke. Investigate before celebrating.
Monthly: refresh evergreen pages, ship a 'what changed' block
We measured a +38% Perplexity citation lift on pages that ship a dated changelog block versus a control set that did not. Refresh dateModified honestly, never rewrite publishedDate.
Quarterly: re-baseline against 1,000 prompts and a competitor set
Once a quarter, expand the sample. Add three competitors. Re-baseline. Anything below 12% citation share in a category where you have product-market fit is a red flag that needs a quarterly content sprint.
Quarterly: kill or rewrite content older than 90 days with zero citations
See the abandonment criteria in the next section. Most GEO programs accumulate dead pages. The number of dead pages a team carries is inversely correlated with overall citation share.
The 4 abandonment criteria (when to kill a page)
Most GEO programs accumulate dead pages. They were good in 2024. They are invisible in 2026. The teams that win are the teams that prune. Here are the four criteria we use.
- Zero citations for 90 days across all six engines. The page failed the witness test. Rewrite the first 80 words of every H2 or unpublish. In our customer set, pages that hit this criterion recover in less than 6% of cases without a structural rewrite.
- Decay greater than 50% of the original peak in 60 days. The model got a fresher source. You are not going to win the topic back without a structural rewrite. Layering in a stat refresh and a single new H3 lifts about 25% of these pages back.
- Topic drift below 10% relevance. A model now cites your page for a question you did not optimize for. Either you re-target the page to the new query or you let the new winner take the slot. Holding a page in the wrong topic costs internal link equity.
- Backlink profile that does not match the page’s claim. No third-party verification = no Claude citation, period. Earn one inbound from a publishable source within 60 days or remove the page from your indexable set.
A before / after we ship on every measurement audit
# BEFORE (vanity dashboard, untraceable)
Citation share (ChatGPT): 38%
Trend: up 4% week-over-week
Top competitor: Acme Inc.
# AFTER (operative dashboard, decision-grade)
Citation share (ChatGPT, /pricing): 38% (+4 pts WoW)
Citation share (ChatGPT, /vs-acme): 12% (-3 pts WoW) <-- regression
Engine spread: 4 / 6 (Claude, Gemini missing)
Passage prominence: 1.6 / 3.0 (mostly mid-answer)
Answer churn (48h): 29% (>= 25% noise floor: monitor)
Action: rewrite first 80 words of /vs-acme; add 1 named source per 150w.
Re-measure in 7 days. Kill criteria: -50% peak by 2026-06-15.The second version takes the same 10 seconds to read but produces a decision. That is the bar. If your current tool cannot generate the second version with one click, it is a monitoring tool, not a measurement stack.
Where teams get GEO measurement wrong
We have audited around 60 SaaS measurement setups in the last six months. Four mistakes account for roughly 80% of the lost signal.
- Counting brand mentions instead of citations. A brand mention is “Notion” showing up in a paragraph. A citation is “Notion” answering the user’s question. Mention-count tools (Meltwater-style PR trackers repurposed for AI) over-report by 4 to 8x against citation-share tools on the same prompt set.
- Tracking only ChatGPT. ChatGPT generates more raw volume than Perplexity and Claude combined, but the conversion math is different. Perplexity’s median visitor is materially more likely to convert on a B2B SaaS demo. Claude cites less often but quotes longer passages, which lifts trust. A ChatGPT-only setup misses both.
- Trusting GA4 + UTM theatre. ChatGPT preserves a
chat.openai.comreferrer roughly 60% of the time. Claude preserves almost none. Perplexity preserves close to 100%. A GA4-only stack undercounts AI traffic and misreports the engine mix. Pair GA4 with server-side bot logs (GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot) and a citation-share tool. - Sampling once and shipping the number. A single 38% citation share measurement is statistically meaningless. Run a fixed prompt set of 200+ items weekly, or do not sample at all.
Where to go deeper
Six companion playbooks walk through each block of the measurement stack in detail, with the prompts, scripts and scorecards we ship to customers.
- How to Measure GEO PerformanceThe operating playbook for your weekly 200-prompt pull.
- GEO Metrics: Share of Voice, Citations & SentimentThe formal vocabulary, with the reference formulas.
- Best GEO Tools Comparison 2026Extended teardown of the 8-tool table above, with raw run logs.
- How to Track AI Search VisibilityThe technical wiring: server logs, bot detection, crawler analytics.
- Competitor Analysis in AI SearchHow to baseline your top 3 competitors before setting targets.
- GEO Reporting & DashboardsWhat to put in front of a CMO every Friday.
Measurement is a sales pitch. Pick your metric carefully. Citation share with formulas, sampling rules and kill criteria sells your CMO on a 12-month roadmap. A vanity number on a dashboard sells her on cancelling the contract.
