How do you measure your brand's visibility in AI search?

What are the two things to measure, discovery and trust?

Measurement answers two separate questions. Discovery asks whether a buyer who does not know you can find you on a question with no brand in it. Trust asks what the engines say when someone checks you by name. The full method is on the page how is a brand's visibility in AI answers measured?

The working steps:

1. Write a fixed set of unbranded questions buyers ask, and a second set that names your brand.
2. Ask every engine every question and store the whole answer: the brands named and the sources cited.
3. Report one row per engine and per question, never one blended score.
4. Repeat on a schedule and read any change against run-to-run noise.
5. Group cited sources by where they originate, not by copies. Ten syndicated copies of one press release are one claim (New Map review of crypto AI-visibility research, October 2026).

The loop below shows how the steps feed each other.

Why must each engine be measured on its own?

Engines behave differently, even on identical questions. In New Map's measurement² (week of 12 October, 68 discovery questions asked once on each of Gemini, ChatGPT and Perplexity, 204 answers), Gemini cited 8.3 sources per answer, ChatGPT 4 and Perplexity 0.3. Only 11 of 68 Perplexity answers cited any source. Method: how New Map measures.

The same week, 30 of 68 Gemini answers named a brand, against 22 of 68 for ChatGPT and 26 of 68 for Perplexity (New Map's measurement). Where an engine rarely shows sources, counting mentions carries the measurement. This is one week and one reading, not a trend.

Source mix also varies by engine. Reddit took 5.4% of citations in Google AI Overviews, 5.3% in AI Mode and 0.24% in ChatGPT (Otterly, 20 September 2026, 1,600 prompts daily on seven engines). Treat each surface as its own row. The chart shows the citation gap in our own sample.

How do you measure share of voice and citations without confusing them?

Count them as two metrics. Share of voice is how often the answer names you. A citation count is how often your pages appear as a source. They diverge: 43% of answers that cited a brand's own event page never named the brand (Ahrefs³, July 2026, 9,886 answers across 34 pages). And 69% of self-promotional listicles cited in AI Overviews belonged to brands the answer did not recommend (Lily Ray, June 2026, 100 queries).

Also record which third-party sources the engines lean on. AirOps reports that third-party signals account for 85% of brand visibility in LLMs (via Kevin Indig, October 2026). In New Map's measurement, reddit.com appeared in 23 of 204 answers (11.3%) and wikipedia.org in 0 of 204. The list of most-cited domains shows where a brand could be present.

How do you measure trust when clicks and referrals are unreliable?

Read the answer itself, because referral data is getting worse. Profound⁴'s observational data shows that 98% of ChatGPT answers with citations carried a tagged source link on 8 October 2026, 75% on 9 October and 66% by 14 October (sample size not stated). Links inside opened panels kept their tags, while links in the first answer often lost them. Traffic reports will undercount.

Clicks were never the main prize. Users clicked a link inside a Google AI summary in 1% of visits (Pew, March 2025, 900 adults). Among US Google searches, the zero-click share went from 60.45% in 2024 to 68.01% in early 2026 (SparkToro and Similarweb, via Kevin Indig, October 2026).

The influence shows up elsewhere. SparkToro and Similarweb followed clickstreams for seven days after an AI engine recommended Capital One: direct visits climbed, searches fell by roughly 15%, and the overall lift came to 14.2% (via Kevin Indig, October 2026). Forrester found 94% of business buyers use AI search when buying (via Ahrefs, October 2026, sample not stated). Our read: measure what the engines say about you by name, and track direct visits beside it.

How much can one reading tell you?

Less than a dashboard suggests. ChatGPT and Google AI repeated the same brand list in under 1 in 100 runs (SparkToro, January 2026, 2,961 runs). Weekly moves under about 5 points can be chance at typical run counts (Peec AI, September 2026, 120M chats).

Citations also fade. The median page lost half its citation share within 11 days of peak (Profound, September 2026, 1.19M page-engine lifecycles). So a single snapshot proves little, and a repeated one compared to a noise band proves more.

New Map's 204-answer reading covers one week and one question set. It says where brands and sources appeared that week. It cannot show whether anything is rising or falling.

What to do

1. Write a fixed set of unbranded buyer questions plus a by-name set, and run both on each engine separately, storing named brands and cited sources (New Map's reading used 68 questions on 3 engines).
2. Report named-brand rate and citation count as two columns for each engine, and flag any weekly change under about 5 points as possibly noise (Peec AI, September 2026).
3. Rank the third-party domains cited on your questions, correct wrong facts at their origin, and compare options on plans and pricing only after your baseline exists.

Frequently asked

Sources

1. Kevin Indig, How to plan 2027 when traffic is down and AI won't attribute, growth-memo.com, 5 October 2026
2. New Map measurement, How we measure, 18 October 2026
3. Ahrefs, AI Search ROI Is Messy. Here’s How to Measure It Anyway, ahrefs.com, 6 October 2026
4. Profound, ChatGPT’s Missing UTM Tags: What Our Research Shows, tryprofound.com, 14 October 2026