How do I track competitors in AI search?

Q: How do I track competitors in AI search?

Track competitors in AI search by re-running your category prompts on a fixed cadence and logging what changes: each brand's share of voice, new competitors that appear, sentiment shifts, and which sources start citing them. A one-time audit is a snapshot; tracking catches the drift, like a rival climbing from 10% to 25% of answers over a quarter before it costs you deals.

Q: What should I track about competitors over time?

Four signals: share of voice for your category prompts, new competitor names that appear month over month, sentiment shifts in how engines describe rivals, and changes in which sources (G2, Reddit, listicles) the engines cite when recommending competitors.

Q: Is a monthly manual sweep good enough to track competitors?

For fewer than 10 prompts where you only care about big swings, yes. But LLM answers are non-deterministic, so a single run per prompt per month is noisy and can mistake a fluke for a shift. Past a handful of prompts, automated tracking that repeats prompts and plots trends is more reliable.

Q: How is tracking competitors different from a one-time AI search audit?

An audit tells you where every brand stands today. Tracking tells you which direction each brand is moving. A competitor doubling its share of voice - say, from 12% to 24% - only shows up when you measure the same prompts repeatedly, not in a single snapshot.

Question

How do I track competitors in AI search?

Maciej Grabek · Accepted Answer

Tracking competitors in AI search means re-running your category prompts on a fixed cadence and logging what changes: each brand's share of voice, new competitors that appear, sentiment shifts, and which sources start citing them. A one-time audit is a snapshot. Tracking catches the drift, like a rival climbing from 10% to 25% of answers over a quarter before it costs you deals. The four signals worth tracking An audit answers "where do I stand today?" Ongoing tracking answers "which way is this moving?" Those are different jobs, and the second one is where competitive decisions come from. Watch four signals over time rather than one headline number. Share of voice for your category prompts. Of all the answers to a prompt like "best CRM for startups", what fraction name each brand. This is the core metric because it is relative: you can hold steady and still lose ground if a competitor doubles. New competitor mentions. Names that appear this month that were absent last month. A brand you have never heard of showing up in ChatGPT and Perplexity answers is an early signal, usually weeks before it shows in your sales calls. Sentiment shifts. Whether an engine's description of a rival gets warmer or cooler. "A solid budget option" turning into "the category leader" is a move worth catching. Source changes. Which pages the engines cite when they recommend a competitor: a new G2 category badge, a fresh Reddit thread, a listicle that added them. Sources are the lever, so knowing which one moved tells you where to respond. Manual sweep vs automated tracking You can do this by hand. Pick 10 to 20 category prompts, run each on ChatGPT, Perplexity, Gemini, Claude, and DeepSeek, and tally which brands each answer names in a spreadsheet. Repeat monthly. It costs nothing but time, and for a handful of prompts it is a reasonable start. The catch is noise. LLM answers are non-deterministic, so a single run of a prompt on a single day can name a competitor that a second run would drop. One data point per prompt per month means you cannot tell a real shift from a fluke, and a manual method quietly drifts as whoever runs it changes wording or skips an engine. Automated tracking runs the same prompts on a schedule across all five engines, repeats each prompt to smooth out the randomness, and plots trend lines instead of single readings. It catches gradual drift that a monthly snapshot misses. The tradeoff is that it needs tooling and a stable prompt set. The honest rule of thumb: if you track fewer than 10 prompts and only care about big swings, a monthly manual sweep is fine. Past that, the hand-tallying breaks down. A share-of-voice snapshot over two months Here is what a tracked view looks like. Suppose you track one category prompt across the five engines for two months and tally how often each brand appears. The numbers below are illustrative, but the shape is typical. "What are the best project management tools for creative agencies?" Brand May share of voice June share of voice Asana 32% 30% Monday.com 28% 27% Notion 12% 24% Your brand 18% 16% The headline in June is not that you slipped two points. It is that Notion jumps from 12% to 24%. If you saw a competitor double like that, the next move is to check which engines drove it and which sources changed - often a new comparison or roundup page the engines have started citing. A single-month audit would have shown you at 16% and moved on. Tracking shows you the trajectory, which is what tells you where to look before the trend hardens. Running the cadence Lock a prompt set. 10 to 25 category prompts, buyers' language, no brand names in the prompt. Do not change them month to month or your trend line is meaningless. Fix the engines and the frequency. Same five engines, same day of the month. Consistency matters more than volume. Log the four signals, not just mentions. Share of voice, new names, sentiment, and cited sources go in the same record so you can read them together. Review the deltas, not the absolutes. The useful column is month-over-month change. Flag any competitor that moves more than a few points and trace it back to a source. Doing this by hand is realistic for one small prompt set. Beyond that, avisibli runs the sweep for you: it tracks share of voice for your category prompts across ChatGPT, Perplexity, Gemini, Claude, and DeepSeek, flags new competitors as they appear, and shows which sources moved the needle, so you watch the trend instead of rebuilding the spreadsheet. avisibli is the GEO platform that publishes this answer library. Self-references are limited to topics where a tool-based answer is genuinely useful to readers.