Competitor monitoring in AI answers

The answer is a set, and the set changes

There is no stable "first place" in an AI answer. In SparkToro's research published in January 2026, 600 volunteers ran 12 recommendation prompts through ChatGPT, Claude and Google's AI—AI Overview, or AI Mode when an Overview did not appear—for a combined 2,961 runs, between November and December 2025. Sometimes the answer contained two or three recommendations; equally often, ten or more. The chance of ChatGPT or Google's AI returning the same list of brands in two runs was below 1 in 100. The authors, in work that is not peer-reviewed, conclude that order within the answer is nearly random and so is not a metric, and that the percentage of presence, measured across dozens or hundreds of prompts run several times, is probably a reasonable one.

The leading brand also changes from model to model. An exploratory preprint on arXiv, in version 2, revised on September 23, 2026, put 250 queries written by the researchers, across five industries, to three models through their APIs, each query five times in two rounds, February and September 2026, which yielded 3,614 and 3,750 scored answers. The three models named the same top brand, among the 50 brands the researchers sampled, in 143 of the 250 queries, or 57.2%, in February, and in 60.8% of them in the September repetition; a query in which a model named none of the sampled brands counts as disagreement. In about four queries in ten, then, the brand that leads on one model is not the one that leads on another.

The consequence for measurement is direct. The competitor has to be observed in the same answers in which your brand is observed: same question, same platform, same run. A collection made on another day, with similar questions, measures a different set. Only paired collection lets you say "in this answer, they got in and I did not," and only repetition lets you say how often that happens.

What only shows up with the competitor in the same denominator

Measuring your own brand produces a number. Measuring your brand and your competitors in the same answers produces four things the number alone does not show.

The first is the actionable gap: the answers in which the competitor appears and your brand does not. That is the subset worth reading first, because each of those answers holds a question the buyer asked, an option the AI offered and your absence, all on the same page. The second is the differentiating criterion: in answers where both brands appear, the AI may separate them by something, team size, price, a feature, an audience, and that criterion is the cheapest hypothesis you will get about how your category is being explained. The third is the set of attributes attached to each brand, which may be correct, outdated or swapped between competitors. The fourth is the change in the composition of the set over time: who came in, who dropped out, and whether the set grew or shrank while your presence stayed the same.

The example that follows is hypothetical. A company sells a platform for creating and selling online courses, Brand A. It tracks 20 category prompts on two AI platforms, with three collections of each prompt in the period, which gives 120 valid answers. Two competitors the sales team knows well, Brand B and Brand C, are tracked in the same answers.

Brand	Answers with a mention (of 120)	Visibility
Brand A (yours)	36	30.0%
Brand B	72	60.0%
Brand C	48	40.0%

On its own, Brand A's 30.0% visibility (36 ÷ 120) says nothing. Next to Brand B's 60.0% (72 ÷ 120), it says the category is answered with Brand B twice as often. The cross-tabulation is what matters: of the 120 answers, 24 mention A and B, 12 mention A without B, 48 mention B without A and 36 mention neither (24 + 12 + 48 + 36 = 120). The 48 answers with B and without A are 40.0% of the period (48 ÷ 120), and they are the gap.

Reading the 48 answers changes the conversation. Suppose 30 of them come from a single topic, six prompts about selling courses on a recurring subscription, which add up to 36 answers in the period (6 prompts × 2 platforms × 3 collections); within that topic, Brand B appears in 33 and Brand A in 3, always alongside B, which leaves 30 answers with B and without A. The gap is not spread across the category: it is concentrated in a subject Brand A may not even cover on its own site. In the 24 answers where both appear, reading shows the criterion: the AI presents Brand A "for people just starting out" and Brand B "for schools with a team." If Brand A does serve schools with a team, that is an attribute to fix in the public information, not an opinion of the AI to regret.

In the following period, with the same 20 prompts, Brand A is still in 36 answers. But a sample read shows a new name, Brand D, a course marketplace, in 40 answers. Nobody registered it, so no number recorded it; Brand A's visibility stayed at 30.0% while the set gained a participant. Registering it makes Brand A's share of voice, its share of the mentions among the compared brands, fall from the following period on, because the set now has four brands and Brand A is still in the same 36 answers. The calculation, and the sample size needed for a change to mean anything, belong to the glossary entry on AI share of voice; the practical consequence is that the number moves with the list, and the list needs to be recorded with a date.

The blind spot of a closed list

A closed list of registered brands is what makes the comparison reproducible from one month to the next, at the cost of not seeing who was left out. Why extracting every brand the AI names does not solve it is covered in the glossary entry on AI share of voice. The fix is not to change method: it is to review the list at the right moments.

The blindness becomes a problem in three situations. When the category is in motion, with new entrants, mergers or new products from large companies. When your visibility stays flat but the answers get longer or start naming more brands; something got in, and it was not you. And when a new topic is added to the prompt set, because new questions bring new sets.

The review is done by a person, and it should be. Once a month, or whenever the prompt set changes, read a sample of the original answers, especially the ones that mention no registered brand, and note the names that recur. A name that appears in one answer is noise; one that appears in several, on more than one platform, is a candidate. When you register it, record the date, because the number is then calculated with a different list.

Which competitors to track

The competitors your sales team faces are not always the competitors the answer presents. The AI responds to the buyer's question, not to the company's competitive map, and "which platform for selling courses" can be answered with a marketplace, a comparison site that lists ten options, a review aggregator, a tool from a neighboring category that "also works" or a suggestion to do without a platform at all. Those options occupy the slot your brand is competing for. An answer competitor is any name the AI offers the buyer in the same answer where your brand should be, even if it does not sell what you sell.

The selection criterion is four questions, in order: does the name actually appear in the original answers to your prompts, and not just in your sales deck? Does it compete for the same buyer decision, or does it appear by chance, as a counterexample or for another need? Can you act on the difference, by correcting an attribute, covering a topic or earning presence in the source the AI cites? And does it appear consistently, on more than one platform and in more than one prompt, or did it appear once? A name that passes all four goes on the list; one that passes only the first is a candidate to watch, not to register.

Tracking too many has a cost, and it is not only the cost of a slot. Each additional brand grows the share of voice denominator and shrinks everyone's share, which makes changes smaller and harder to read. Each additional brand is also one more subset of answers to read. Five competitors chosen by importance say more than fifteen chosen for completeness, because the gap that matters is the gap against whoever wins the answer often. Start with two or three names taken from reading the answers, not from the sales spreadsheet, and reserve the remaining slots for what the monthly review reveals.

What competitor monitoring has to deliver

The general requirements of monitoring, a stable prompt set, collection outside your browser, the original answer preserved and an explicit denominator, are on the page about GEO monitoring. Competitor comparison adds requirements of its own, and each can be checked in a trial account before deciding on any product.

Requirement	Why it matters	How to check
Competitor measured in the same answer	Separate collection measures a different set; comparison needs the same question, platform and run	Open one answer and see every detected brand in it, together
One presence per brand per answer	Counting every occurrence of the name favors long answers and inflates whoever is repeated	Ask for the counting rule in writing
The gap as a list of answers, not a number	What you read and fix is the answer; the percentage only points where to look	Filter the answers where the competitor appears and you do not, and open one
Filters by platform, topic and prompt	A gap concentrated in one topic calls for one action; a scattered gap calls for another	Check that the same filter applies to the gap and to the total
Brand list dated and kept with each answer	Without it, you cannot tell which list produced last month's number	Ask which list produced each number and whether the original reading is kept after a change
Export with mentioned brands per row	Lets you cross presence by brand with topic, platform and date outside the dashboard	Export a period and look for the brands column
A statement of what the metric does not say	A mention is not a recommendation or sentiment; a dashboard that does not say so misleads	Ask for the definition of a mention and what it excludes

A sentiment score or a "recommended" badge stays off the list because each is a separate measurement, an automatic classification of the text, and not a requirement of competitor comparison. Whoever offers one should show the original answer that supports it.

How Lastfind makes the comparison

In Lastfind, a project has your brand and up to ten competitors. Collection depends on where Lastfind runs: the free Cloud account collects each prompt once a week, on Mondays at 4 AM Brasília, on ChatGPT only, and an independent installation collects daily on your own infrastructure, on the channels your provider account allows. Each answer records which of the registered brands were mentioned. A mention is the detection of the brand name in the final answer text, with word boundaries and excluding link addresses; the criterion the answer uses to separate the brands, like the attributes attached to each one, is read in the original answer, not classified by the mention count. Sentiment evaluation and semantic branding are separate, optional features, off by default and outside this comparison; sentiment reads your brand and the first five competitors, while the others have no sentiment reading.

The project summary shows a ranking with your brand and the tracked competitors, the visibility of each, the share of voice, which is the share of mentions among the registered brands counting each brand once per answer, and the comparison with the preceding period of equal length. The analysis breaks the result down by channel, topic and prompt and lists the opportunities: the answers that mention another tracked brand and do not mention yours. That is the gap from the example above, filterable and openable, because each answer can be opened and its original text and sources read in full.

History is paginated, and the CSV export covers the whole filtered period, with the mentioned brands on each row, for anyone who prefers to cross the data outside the dashboard. Reports cover 1 to 90 days; older data stays stored in the project.

The original text of each answer and the collection settings are preserved. When you change the tracked brands or competitors, Lastfind recalculates the mentions in the stored answers of the last 90 days (up to the 10,000 newest) without requesting them again: the report follows the current list and the original collection record is kept, so record the date of each change. The ranking and the share of voice compare the brands you registered, not the whole market, and a brand you did not register appears in no number. Lastfind suggests names found in comparable answer contexts, each with its evidence, and you choose which ones to track; reading a sample of the original answers remains the way to check them.

When Lastfind is not the right tool

Lastfind does not monitor Copilot or Grok, because the collection provider documents no endpoint for them, does not generate automatic recommendations about what to do with the gap and has no public API or outbound integration with Slack, spreadsheets, Zapier or alert webhooks. If your core need is one of these, another tool will serve you better.

First step

Write ten category questions, without your brand name, in the language the buyer asks in, and register your brand and the two or three names you believe occupy the answer. Lastfind's free Cloud account, with one topic, ten prompts and ChatGPT only, runs those questions every Monday at 4 AM Brasília, with no payment details required, and gives you a ranking, opportunities and the original answer; new prompts enter the next Monday window. After two Monday collections, open the answers in which no registered brand appears and read who is there. The initial hypothesis then gives way to the list of brands observed in those ten questions, on ChatGPT, during that period, not to a general list from the AI. Anyone who needs other channels or daily collection can install their own version, which is AGPL-3.0-only and runs on their own infrastructure with their own provider credentials; there, Gemini requires an English-language project, and completion depends on each provider, with up to 72 hours on Claude.

Platform documentation checked in October 2026.