There is no official source of competitive data on AI Mode. Not in Search Console, not in Analytics, not through an API. What Google publishes about generative surfaces concerns your own site and nothing else. Comparing your position to a competitor's therefore means building that comparison yourself, by observing what the engine answers.
This article covers what the tools genuinely give, the protocol that makes the comparison defensible, the five indicators to record, and the four mistakes that invalidate a reading.
Can you see your competitors in AI Mode using Google's tools?
No, and this is not a temporary limitation of the rollout. Search Console is a property tool: it reports what concerns the domain you have verified. It has never shown a third party's data, and the generative AI report is no exception.
So no Google tool answers the question "does my competitor appear more often than I do in AI Mode". Third-party tools that advertise it are not reading Google's data: they do exactly what you can do yourself, which is ask the surface questions and record the answers. That is worth knowing before buying a subscription, because it moves the question from "who has access to the data" to "who has the more rigorous protocol".
What does Search Console's generative AI report actually show?
Impressions of your pages in Google Search's generative AI features, broken down by page, country, device and date. Google announced it on 3 June 2026, tested it on a subset of properties, then rolled it out to all properties worldwide on 31 August 2026. Its blog post states that the reports give dedicated views of impressions within generative AI features on Search, such as AI Overviews and AI Mode, as well as in Discover, and that this data continues to be counted in the overall performance report.
Four absences matter more than what is present.
| What the report gives | What it does not give |
|---|---|
| Impressions of your pages | Clicks and click-through rate |
| Breakdown by page, country, device, date | The queries behind the appearance |
| Separate Search and Discover views | Any split between AI Overviews and AI Mode |
| Your own domain | Any competing domain |
Source: Google Search Central, "Introducing Search Generative AI performance reports in Search Console", 3 June 2026, and related documentation. The worldwide rollout has been effective since 31 August 2026; some properties still lack access, and the report only shows data above a sufficient impression volume.
The third line is the one most often forgotten. AI Overviews and AI Mode are two different surfaces with different citation behaviour, and the report counts them together. A rise in your generative impressions can therefore come entirely from AI Overviews while your presence in AI Mode collapses. The number goes up, the reality goes down.
What is the only method available then?
Observe AI Mode's answers directly across a fixed panel of questions, and record who is named. This is sampling work, not data extraction. The protocol rests on five conditions, and removing any one of them makes the comparison uninterpretable.
- A fixed panel of questions, settled before the first reading and never adjusted afterwards. The questions must be the ones your buyers actually ask at the moment of deciding, not keywords.
- A named list of competitors, also settled in advance. Without it, you will count the brands that catch your eye rather than the ones that recur.
- Several runs per question, because a generative answer is not deterministic.
- A recorded reading date, with raw answers kept, not just the aggregated score.
- A constant calculation formula from one reading to the next.
A question asked twice in AI Mode produces two answers, sometimes with different brands. A single reading therefore measures one draw, not a position. Before concluding that a competitor is ahead of you, you must have measured the band of variation on a single question — otherwise the observed gap can be entirely noise.
Which indicators should you compare?
Five are enough, and all of them are read from the visible answer.
Presence rate is the share of panel questions where the brand is named. It is the simplest indicator to compare across competitors, and the most stable.
Rank of appearance is where the brand sits in the answer. Being named first and being named last do not have the same effect on a buyer reading three names.
Share of voice divides your mentions by the total brand mentions across the panel. It is the one measure that literally answers "am I ahead or behind", and it is computed without knowing anything about your competitors' own data.
Cited sources supporting each brand. This is the most actionable of the five, because it shows what a competitor's presence rests on. If three of its citations come from an industry directory you are not listed in, you have found a concrete action.
Sentiment attached to the mention. Being named as the cheap option and being named as the sector reference do not sell the same way.
Presence rate and share of voice compare directly. The other three are read, not added up.
Does visibility in AI Mode say anything about other surfaces?
Very little, and that has been measured. AI Mode and AI Overviews belong to the same engine and still cite largely different sources. A reading taken on one does not describe the other.
The gap is wider still between engines. Our own readings put numbers on it: the large majority of cited sources appear on one engine only. In other words, a solid presence in AI Mode lets you infer nothing about your presence in ChatGPT or Perplexity, and the reverse holds too.
The practical consequence for a dashboard: one column per surface, never a blended score. A single indicator merging AI Mode, AI Overviews and ChatGPT can rise while the only surface your buyers use goes dark.
The four mistakes that invalidate a reading
Reading from a signed-in account. AI Mode takes user context into account. A reading run from the founder's account, which visits the company's own site every week, is not a neutral observation. Readings are taken in a clean session, under the same conditions every time.
Ignoring location and language. The same question asked in French from Paris and in English from Bangkok does not pull the same sources. If your buyers sit in two markets, you need two panels, not one translated panel.
Changing the panel mid-course. Adding three questions you perform well on between two readings lifts the score without anything having changed. The panel is frozen before the first reading, and additions go into a separate panel.
Automating extraction. Google publishes no API for AI Mode, and automated extraction of its results pages breaches its terms of service. A defensible protocol therefore rests on manual or assisted reading at controlled volume. It is slower than a dashboard that fills itself, and it is the only form that survives a client, a competitor or a regulator asking how the number was produced.
How to read a gap between two brands
A raw gap is not a result until it has been measured against the panel's natural variation. This is the rule we apply to our own readings, and the one we ask of any study we cite.
Concretely: if the same question, asked five times on the same day, names your competitor three times out of five and you twice out of five, the 20-point gap is not a position, it is the amplitude of the noise. You establish that amplitude across the panel first, then treat as real only the gaps that exceed it.
This discipline costs time and it makes slides disappear. It also has the advantage of never selling a programme on a variation that would have corrected itself.
Where this measurement sits in the journey
Comparing your presence to competitors' in AI Mode measures one stage of the path from existing online to being recommended. The KAIS AI Visibility Chain counts seven.
Be Discovered → Be Retrieved → Be Understood → Be Trusted → Be Cited → Influence the Answer → Be Recommended.
A share-of-voice reading answers the "Be Recommended" stage. It does not say why you are there or not. The cited-sources column is what links back to the upstream stages: if your pages are never cited while your name appears, the problem is on your pages; if your name appears nowhere, the problem sits higher up, and the twenty articles you were about to publish will not change it.
Where to start this week
- Write twenty questions your buyers ask before choosing a provider like you. Full sentences, not keywords.
- Name the six competitors you want to track.
- Ask each question three times in AI Mode, in a clean session, on the same day, and record the brands named and the sources cited.
- Calculate your presence rate and share of voice, and the variation from one run to the next.
- Repeat in thirty days with exactly the same panel.
Step four produces the baseline. Without it, the second reading compares to nothing, and in a month you will discover that you measured twice without being able to conclude.
Sources
- Google Search Central — Introducing Search Generative AI performance reports in Search Console, 3 June 2026. Dedicated views of impressions within generative AI features on Search, including AI Overviews and AI Mode, and in Discover; data also counted in the overall performance report.
- Google Search Central — AI features and your website. Eligibility conditions for generative features.