Question

How to Build a Repeatable AI Search Visibility Benchmark

By Jarno Saarimies, Founder of AEOvara · Last updated: September 8, 2026 · 9 min read

Your buyers are already asking ChatGPT, Perplexity, and Google AI Overviews who to trust. Google says AI Overviews now reach 2 billion monthly users, and industry tracking puts them on roughly 48% of all queries — up 58% year over year (BrightEdge, February 2026). Meanwhile, Seer Interactive's longitudinal study found organic CTR on AI Overview queries dropped from 1.76% to 0.61% between June 2024 and September 2025.

Here's the problem: your analytics stack is blind to all of it. If ChatGPT recommends a competitor and never mentions you, nothing gets logged. No click, no impression, no bounce rate. The opportunity just evaporates.

A benchmark fixes this. Not a one-off audit you run once and forget — a repeatable measurement system that tells you, month after month, whether you're becoming more or less visible inside AI answers.

This guide shows you exactly how to build one, starting with a free spreadsheet method and scaling to automated tracking when the data demands it.

What "AI search visibility" actually means (define it before you measure it)

Before you measure anything, agree on what you're measuring. Five metrics carry the weight:

1. Mention rate — the percentage of tested prompts where your brand appears at all. This is your headline number.
2. Share of Voice (SoV) — your mentions as a percentage of all tracked-brand mentions in the same answers. Peec AI's own documentation shows why this matters: a brand mentioned 4 times against a competitor's 12 mentions has high visibility but only 25% SoV. Visibility is presence. SoV is competitive prominence.
3. Citation rate — whether the answer actually links to you, not just names you. Track this separately: on ChatGPT most mentions carry no link, while Perplexity links almost everything.
4. Position and prominence — where you land in the answer. AI engines tend to treat the first-named brand as the default recommendation. Buried in the last paragraph is not the same as leading the list.
5. Sentiment and accuracy — not just whether you appear, but how you're described. A confident but outdated mention (wrong pricing, wrong positioning) can cost you the deal.

Step 1: Build a frozen prompt corpus

Pick 10–15 prompts that real buyers would actually ask. Not keywords — questions. Organize them into five categories:

Prompt type	What it reveals	Template
Category / discovery	Whether you're in the consideration set at all	"What are the best [category] for [use case]?"
Comparison	Positioning against a named rival	"How does [your brand] compare to [competitor]?"
Persona / use-case	Whether you're recommended in your buyer's context	"What should a [role] look for when choosing a [category]?"
Problem / jobs-to-be-done	Top-of-funnel presence before your category is named	"How do I [problem]?"
Brand-direct / reputation	Accuracy and sentiment	"What is [your brand]?" / "Is [your brand] any good?"

This list stays mostly frozen. This matters more than most people realize. If you change the prompts every month, you're not tracking visibility — you're running random tests. You can add prompts once a quarter, but the core set must stay stable or your trend data is meaningless.

Step 2: Choose your engines and cadence

Measure each platform separately, because a strong ChatGPT score tells you nothing about Perplexity or AI Overviews. At minimum: ChatGPT, Perplexity, Gemini, and Google Search with AI Overviews.

On cadence: weekly tracking is the standard for most teams (LLM Pulse and similar platforms default to this). Daily tracking makes sense for product launches or volatile categories — finance shows the highest citation volatility of any vertical. Monthly is the absolute floor for a manual process.

Run each prompt 3–5 times per engine. Generative answers are probabilistic — the same question names different brands on different runs. One query is an anecdote, not a measurement. Research on robust benchmarks suggests 500–2,000 prompts per week for statistically stable values (SUMAX Research, 2026); for a manual baseline, 10–15 prompts × multiple runs × 3–4 engines is a realistic starting point.

Step 3: Run the manual baseline (this afternoon, free)

You don't need a platform yet. You need a spreadsheet and about an hour.

Run your prompt set through each engine — same wording, same account settings where possible. For each response, log:

• Query text
• Platform and date
• Mentioned? (Yes/No)
• Position (1st named, 2nd, 3rd…)
• Cited with a link? (Yes/No)
• Competitors named
• Sources cited in the answer
• Accuracy/sentiment notes
• Screenshot

Screenshot everything. AI results move around. Something that shows up this month may disappear next month. If you want to understand trends, you need receipts.

The last column is usually the most useful. If Perplexity or another engine shows sources, those sources tell you where the answer gets its confidence. Sometimes the fix isn't "write another blog post" — it's getting your company described better on the third-party pages the AI is already reading.

Step 4: Calculate your benchmark numbers

Four formulas turn raw logs into a benchmark:

• Mention rate = prompts where you appear ÷ total prompts run × 100
• Citation rate = prompts with a link to you ÷ total prompts × 100
• Share of Voice = your mentions ÷ all tracked-brand mentions × 100
• Position score = average position when mentioned (lower is better)

Example: mentioned in 5 of 12 prompts on Perplexity, with 4 of those including a link. Mention rate: 42%. Citation rate: 33%. Your competitor mentioned in 9 of 12 with 7 links. SoV: 5 ÷ 14 = 36%.

That gives you a real baseline — and usually a few uncomfortable surprises. One thing to internalize early: position in a single answer is mostly noise; frequency across many runs is the signal (as SparkToro's analysis showed). Track how often you appear, not where you landed once.

Step 5: Add context metrics the spreadsheet can't see

A prompt log tells you that you're visible or invisible. Three additional signals tell you whether it matters:

AI referral traffic. Expect attribution to undercount. Add a "How did you hear about us?" option with an AI assistant choice to your intake forms, because most AI mentions never produce a trackable click.

Branded search lift. Growth in people searching your company name often correlates with AI exposure, per Percepture's GEO research.

Conversion quality. The traffic you do get is disproportionately valuable: Ahrefs' own data showed AI-referred visitors converting at 23x the rate of traditional organic visitors (0.5% of traffic generated 12.1% of signups, June 2025). Semrush found a similar 4.4x pattern. Small numbers, high intent.

Step 6: Automate when manual stops scaling

Twenty-five prompts run five times each is 125 data points — enough for a baseline, not enough for stable rates. That's when a platform earns its money. Roughly priced landscape as of mid-2026:

• Otterly AI — from $29/month (15 prompts). The low-cost serious entry point; weekly refresh.
• Peec AI — from $95/month (50 prompts). Clean visibility, position, sentiment, and SoV dashboards; strong for structured reporting.
• Semrush AI Visibility Toolkit — from $99/month standalone. Low friction if you're already on Semrush.
• Ahrefs Brand Radar — from $199/month per AI-platform index (requires an Ahrefs base plan from $129/month). Prompt methodology anchored to search data.
• Profound — enterprise positioning, deep analytics, most advanced features behind custom pricing.

Whichever tool you pick, demand these capabilities: multi-platform tracking (ChatGPT, Perplexity, Gemini, AI Overviews, Claude), citation/source URL analysis — not just mention counts — competitor benchmarking, historical trends, and transparency about prompt volume methodology.

The seven mistakes that break benchmarks

1. Chasing rank position instead of visibility rate. Position swings wildly run to run; frequency is the stable metric.
2. Measuring only one engine. You're optimizing blind.
3. Counting mentions but ignoring citations (or vice versa). They're different metrics doing different jobs.
4. Ignoring sentiment and accuracy. Being named with wrong pricing isn't a win.
5. Running each prompt once. Anecdote, not measurement.
6. Expecting clean click attribution. Analytics will always undercount AI's influence.
7. Testing once and moving on. Visibility drifts as models update and content ages. Only a fixed cadence reveals what's actually moving.

The bottom line

You can't manage what you can't see, and traditional analytics can't see AI answers. The process is unglamorous: freeze a prompt set, run it repeatedly across engines, log mentions and citations separately, review monthly, automate when the volume outgrows your spreadsheet. Do that consistently and "how visible are we in AI?" stops being a worry and becomes a number you can move.

FAQ

What is an AI search visibility benchmark?
An AI search visibility benchmark is a fixed set of buyer-relevant prompts that you run repeatedly across AI search engines (ChatGPT, Perplexity, Gemini, Google AI Overviews) to measure how often and how accurately your brand appears in AI-generated answers over time. It turns "are we visible in AI?" from a vague worry into a trackable number with a trend line.

How many prompts do I need for a reliable benchmark?
For a manual baseline, 10–15 prompts across five categories (category, comparison, persona, problem, brand-direct), run 3–5 times per engine, gives directional data. For statistically stable visibility rates, research suggests 500–2,000 prompts per week — which is where automated platforms become necessary.

What's the difference between mentions and citations?
A mention means the AI names your brand in the answer text. A citation means it links to your website as a source. Mentions build brand association; citations drive trackable traffic. On ChatGPT most mentions carry no link, while Perplexity links almost everything, so the two metrics must be tracked separately.

How often should I re-run my benchmark?
Weekly is the standard cadence for automated tracking; monthly is the minimum for manual runs. Re-run the exact same prompt set each time — changing prompts between runs destroys comparability. Also re-baseline after major model updates, since visibility can shift overnight when models change.

Can I build an AI visibility benchmark for free?
Yes. A spreadsheet, 10–15 frozen prompts, and one hour a month covers the fundamentals — mention rate, citation rate, position, competitors, and cited sources. Paid tools earn their cost when you need scale, daily cadence, sentiment scoring, and source-level analysis across hundreds of prompts.

Do I still need traditional SEO tracking?
Yes. Ranking on page one remains correlated with AI citations — Ahrefs found 38% of AI Overview-cited pages also rank in the top 10 — but the relationship is weakening fast (down from ~76% in mid-2024). The two systems measure different things: blue-link position versus presence inside the answer. You need both.

---

Jarno Saarimies is the founder of AEOvara, a Finnish consultancy specializing in AI search optimization (AEO/GEO/LLMO) — helping brands get found, cited, and recommended inside AI-generated answers. A professional photographer since 2013, Jarno runs Kuvaajankulma, a photography studio in Lappeenranta, Finland, and applies the same evidence-first mindset to search: no hype, no guesswork — measure what actually matters. AEOvara Jarno S
AEOvara BLOG