First published: 21 June 2026 · Last updated: 21 June 2026
The Three Core Metrics
Every AI search visibility programme needs to track three metrics. Anything beyond the three is either a derivative or vanity. The three are necessary and sufficient.Metric 1: Citation Rate
Citation rate is the percentage of tracked prompts where the brand is cited at least once in the AI engine's response. Formula: `cited_prompts / total_prompts × 100`. Citation rate is the absolute visibility number. A brand with 12% citation rate on a 50-prompt test set is cited on 6 of 50 prompts. A brand with 40% citation rate is cited on 20 of 50. The metric is per engine: a brand may have 35% on Perplexity and 8% on ChatGPT because the engines source differently. Benchmarks from our SG portfolio (n=18 active AEO programmes):- New programme, month 1: typically 0-8% citation rate per engine.
- Programme with Pillar 1-3 audit fixes shipped, month 3: typically 10-20%.
- Mature programme, month 9-12: typically 25-45% on the best-performing engine.
- Best-in-class brands (industry leaders with strong entity signals): 50%+ on Perplexity, 30%+ on ChatGPT.
Metric 2: Citation Share
Citation share measures competitive position. Run the same prompt set against the brand and the top 3 competitors. For each prompt, count which brands are cited. Citation share = `your_citations / (your_citations + competitor_citations) × 100`. Citation share is more strategically informative than citation rate because AI search is a zero-sum competition at the prompt level. A 35% citation rate is excellent if competitors are at 5%; it is marginal if competitors are at 60%. The competitive frame surfaces both the gap and the opportunity. Benchmarks:- Citation share under 10%: brand is invisible to AI search relative to competitors.
- 10-25%: brand is a minor player; competitors dominate.
- 25-50%: brand is competitive; AEO work is paying off.
- Over 50%: brand is the market leader for AI citations on this prompt set.
Metric 3: Citation Rank
Citation rank is the average position of the brand when it appears in a cited-source list. AI engines display 3-10 cited sources per response (varies by engine and query type), and the position matters: in expandable citation UIs (Perplexity's source pills, ChatGPT's footnotes), the first 1-2 positions get the disproportionate click share. Formula: `sum(position when cited) / count(cited prompts)`. Lower is better. Benchmarks:- Average rank 1-2: brand is the canonical source on most cited prompts.
- 2-3: strong, regularly first or second.
- 3-5: present but secondary.
- 5+: cited but not preferred.
Paid Tooling: The 2026 Landscape
The AI visibility tooling category has consolidated around five major players plus a long tail of newer entrants. The five worth knowing:The honest commentary on each:
Profound is the enterprise standard. The prompt-volume feature is genuinely unique: instead of just measuring whether you appear for a prompt, Profound estimates how many users are running that prompt across AI platforms, which converts visibility tracking into demand intelligence. Pricing is opaque and high (custom enterprise contracts typically start in the low thousands per month). Worth it for clients above $50k MRR with serious AEO programmes; overkill for smaller accounts.
AthenaHQ is our default recommendation for clients in the $5k-25k MRR range. The pricing is mid-market, the engine coverage is broad, and the research output (the State of AI Search reports) doubles as content marketing for the customer's own thought leadership. Build quality is high; the founding team's Google Search background shows in the product depth.
Otterly is the price leader and the right answer for clients under $5k MRR or for agencies running their own internal tracking on a budget. The $29/mo entry tier is sufficient for a 20-prompt tracked set across the three biggest engines. The trade-off is shallower analytics depth versus AthenaHQ or Profound.
Peec AI sits in the middle and is a fine choice for accounts that specifically need strong Gemini coverage (which Otterly and AthenaHQ have but Peec emphasises). Less differentiated otherwise.
Brandlight is more of a brand-monitoring tool than a pure visibility tracker. Worth considering for brands where AI sentiment matters as much as citation count (consumer brands, regulated industries, reputation-sensitive verticals).
The honest agency math: if you are running AEO for under five clients, the tooling cost is justifiable; for one or two clients, DIY is more economic for the agency unless the client specifically wants a third-party tool report.
The DIY Methodology (Costs Nothing, Works Fine Up to ~50 Prompts)
The DIY methodology produces defensible numbers and is appropriate for any agency under five active AEO clients or any client where the budget does not justify $29-2000/mo in tooling. The setup is a spreadsheet plus a monthly manual prompt-test cycle.
Step 1: Build the Tracked Prompt Set
The prompt set should be 30-50 prompts representative of the customer's actual queries. Sources for prompts:
- Search Console queries. Pull the top 100 queries the site already ranks for. Filter to questions and informational queries (skip branded, skip transactional).
- Customer interview transcripts. If the client does sales calls, the discovery questions customers ask are gold. Convert each into a prompt form ("how do I [problem]", "what is the best [category] for [use case]").
- Competitor content gaps. Look at competitor article titles in question form; assume the underlying customer question is real.
- AlsoAsked, AnswerThePublic. Mine question-form variants of the primary keyword set.
Convert each query into a natural prompt form. "best crm singapore" becomes "what's the best CRM for a small business in Singapore?". The AI engines respond to natural language better than they respond to keyword strings.
Step 2: Run the Prompts Across Engines
Manually query each prompt in:
- ChatGPT (with web search enabled, GPT-4 or GPT-4-turbo recommended for consistency).
- Perplexity (default mode).
- Google AI Overviews (logged into a Google account, no Personal results filter active).
- Microsoft Copilot (Bing-based).
For each engine and each prompt, record:
- Citation present (yes/no): is the brand cited?
- Citation rank (position 1-N): if cited, what position in the source list?
- Cited URL: which page on the brand's site is cited?
- Competitor citations: which competitors are cited and at what rank?
Time cost: roughly 5-7 minutes per prompt across four engines. A 30-prompt set takes 2.5-3.5 hours per monthly cycle. A 50-prompt set takes 4-6 hours.
Step 3: Compute the Three Core Metrics
In the spreadsheet, derive citation rate, citation share, and average citation rank per engine and overall. Compare to last month's run for trend.
Build prompt set
30-50 prompts from GSC, sales calls, AlsoAsked. One-time setup, refresh quarterly.
Run across engines
ChatGPT (with search), Perplexity, AI Overviews, Copilot. Record citation present, rank, URL, competitors.
Compute metrics
Citation rate, share, rank per engine and overall. Compare to last cycle.
Report and act
Surface losing prompts (cited last month, not this month) for diagnosis. Surface winning prompts as proof of programme working.
Step 4: Bias Controls
DIY measurement has known biases that must be controlled for to produce defensible numbers:
- Personalisation: Use a clean browser profile or incognito with no logged-in account context. Personalised results contaminate the measurement.
- Geography: Use a SG IP (or your client's primary geography) consistently. Switching between SG and US gives different results, especially in AI Overviews.
- Time-of-day: Run at the same time of day each month. AI engines do change responses across time as their indexes refresh.
- Engine version: Note the engine version (GPT-4 vs GPT-4-turbo, Perplexity Sonar vs default). Versions matter and change.
- Prompt phrasing: Once you commit to a prompt phrasing, do not change it month-to-month. Phrasing changes the response. The prompt set is the constant; the response is the measurement.
These controls are the difference between a credible monthly measurement and noise. The paid tools automate the controls; DIY requires discipline.
When to Graduate from DIY to Paid
The economics break around three thresholds:
- Prompt set exceeds 50 prompts. Manual labour at 50+ prompts per cycle exceeds 6 hours per month per client. Paid tooling at $99/mo (Otterly Pro, similar tier) is cheaper than agency time at any reasonable rate.
- More than three active AEO clients. Multiplying manual work across clients consumes a senior strategist's day per week, which is the wrong allocation. Tooling consolidates the workflow.
- Client demands third-party-validated reporting. Some enterprise clients want the reporting to come from a recognised tool brand rather than the agency's spreadsheet. Cosmetic but real.
Below these thresholds, DIY is fine. Above, paid is more economic.
What to Track Beyond the Three Core Metrics
The three core metrics (rate, share, rank) are necessary and sufficient for primary reporting. Useful supplementary metrics:
- Cited URL distribution: which pages on the site are doing the citation work? Often the answer is "two or three pages get 70% of citations", which surfaces the AEO content patterns that work.
- AI referral traffic in GA4: users who click through to the site from AI engine citations. This is the conversion-side metric. Track via GA4 referral source segmentation (perplexity.ai, chat.openai.com, copilot.microsoft.com).
- Sentiment of citation context: how is the brand framed when cited? Positive ("X is the leading provider of Y"), neutral ("X offers Y"), or negative ("X has been criticised for Y"). Brandlight and similar tools surface this; DIY requires manual review.
- Prompt difficulty: which prompts are uncited by anyone (greenfield opportunity) versus heavily contested (multi-citation density)?
These are appendix-level metrics. They support the primary reporting but should not displace it. Our existing ChatGPT citation guide and Perplexity SEO guide cover the engine-specific tactics that produce the metric improvements.
A Worked Example: SG B2B SaaS, 6-Month GEO Measurement Trajectory
Concrete example. Client: SG B2B SaaS, started AEO programme January 2026, 6-month measurement trajectory below. DIY methodology, 35-prompt tracked set, manual monthly cycles.
Observations:
- All three metrics moved together. Rate, share, and rank improved in the same direction every month, which is the signature of a healthy programme. If rate improves but share is flat, you are growing in absolute terms but not relative to competitors (everyone is improving). If share improves but rank does not, you are being cited but not preferred.
- Perplexity moved fastest. Consistent with the engine's preference for fresh, structured, well-cited sources, which the AEO programme delivered.
- AI Overviews moved slowest. Google's AI Overview source selection is more conservative; the lag is normal.
- 6-month trajectory is realistic. 0-4% baseline to 20-37% citation rate is the right shape for a competently-run programme on a B2B SaaS site with reasonable starting authority.
Reporting the Numbers to Clients
The reporting format we use:
- Executive line: "AI search citation share grew from 4% to 28% over six months across ChatGPT, Perplexity, and AI Overviews on a tracked set of 35 high-intent prompts. The brand is now cited on 28% of prompts where competitors are cited, up from 4% in January."
- Per-engine breakdown: the table above with monthly trend.
- Top winning prompts: the 5-10 prompts where citations grew most. Show which content drove it.
- Top losing prompts: the 3-5 prompts where citations were lost or never gained. Diagnose and feed into next month's content roadmap.
- Competitive snapshot: which competitor still leads on which prompt subset.
This format aligns with the broader SEO reporting framework in our SEO reporting template post: business outcome → search outcome → work log. The AI visibility data sits in the search-outcomes layer.
Frequently Asked Questions
Why measure citations and not just AI referral traffic from GA4?
GA4 referral traffic from AI engines undercounts. Many AI users read the cited answer in the AI engine and never click through. Perplexity's source-pill UI shows source URLs but most users do not click them; the answer is consumed in the chat interface. ChatGPT shows footnotes that are similarly low-click. Citation count is the leading indicator (you appeared as a source); referral traffic is the lagging indicator (someone clicked through). Both matter, but citation count is more actionable because it directly reflects whether the AEO work is producing visibility.
How big should the tracked prompt set be?
30 prompts is the floor for a meaningful baseline; 50 is comfortable; 100+ is the right size for enterprise programmes with broad query coverage. Below 30, the noise-to-signal ratio degrades and monthly trend movements get drowned out by a single prompt going wrong. Above 100, manual labour becomes uneconomic and paid tooling pays for itself.
How often should I re-run the prompt-test cycle?
Monthly is the standard cadence and matches typical SEO reporting cycles. Weekly is excessive and noisy; the underlying citation data does not move that fast. Quarterly is too sparse to catch programme regressions early. Pair monthly cycles with a quarterly prompt-set refresh (add 5-10 new prompts, retire 2-3 stale ones).
Are AI engines stable enough that month-over-month comparisons are valid?
Mostly yes, with caveats. ChatGPT and Perplexity engine versions change quarterly and can shift response patterns; note the engine version with each measurement. Google AI Overviews algorithm updates (we have seen four since launch) shift source selection significantly; expect occasional volatility. The remedy is to track the engine version alongside the citation data and to interpret big swings against the announced engine changes. Long-term trends across 90+ days are more reliable than month-to-month deltas.
What about Gemini, Grok, Claude, and other engines I have not mentioned?
Add them if they are relevant to the client's audience. Gemini is increasingly important for the Google ecosystem (Workspace integration). Claude is sourced into Anthropic's product and Claude Code increasingly for technical/dev queries. Grok has measurable share in some verticals. The four-engine baseline (ChatGPT, Perplexity, AI Overviews, Copilot) covers 80%+ of consumer AI search; adding others gets the remaining 20% and matters most in technical or developer-focused B2B verticals.
How do I handle prompts where the brand is not cited but a competitor is?
These are the highest-value prompts to study. Read the full AI response, identify which competitor is cited, visit the cited URL on the competitor site, and identify what makes that page citation-worthy: structure, schema, entity signals, recency, depth. Apply the pattern to your equivalent page. Re-test in the next monthly cycle. Most measurable AEO improvements come from this loop.