How to track competitors in AI answers (and steal their spots)

Competitive intelligence used to mean watching rankings, ads and review counts. In 2026 the shortlist itself moved: Forrester's survey of 18,000 business buyers found 55 percent compared vendors inside AI tools during their most recent purchase, and G2 found 33 percent of B2B software buyers bought from a brand they had never heard of before an AI named it. The commercial weight behind those answers is measurable in traffic too: Adobe Analytics tracked AI-driven referral traffic to U.S. retail sites growing more than 4,000 percent year over year in 2025. When an assistant names three brands for a buying question and yours is absent, a competitor is collecting a lead you never saw exist.

The good news is symmetry. AI answers are composed from readable sources, which means every competitor mention has a paper trail you can follow, and every spot they hold is a spot that can be contested with better evidence. This playbook covers both halves: measuring where rivals stand, then taking the questions worth taking. If the underlying concepts are new, start with what is AI visibility and come back for the competitive layer.

Three numbers carry the whole discipline. Mention rate: the share of repeated runs, per question and engine, in which a brand is named. AI share of voice: a brand's mentions as a percentage of all brand mentions across the tracked question set, the number that turns "how visible are we?" into "how visible are we relative to them?". Citation share: which domains the engines cite when composing those answers, the leading indicator the other two lag. A rival's rising citation share this month is their rising share of voice next month.

Two rules keep the numbers honest. First, unbranded questions only: "best CRM for a small agency" measures the market, "is Acme good?" measures grammar, since a question containing a brand name returns that brand by construction (the full argument is in branded vs unbranded prompts). Second, repeated runs, because SparkToro measured under a 1 percent chance that two identical ChatGPT runs return the same brand list, so a single check is an anecdote about a coin. The deeper math and cadence live in AI share of voice.

Counterintuitive but important: competitor tracking starts with questions, and the competitors emerge from the answers. Draft 20 to 40 unbranded buying questions across the funnel: category questions ("best AI visibility tools for startups"), problem questions ("how do I know if ChatGPT recommends my brand"), comparison questions ("affordable alternatives to enterprise GEO platforms"), and use-case questions phrased the way your actual buyers phrase them on sales calls. Then let the engines tell you who competes: the brands that keep appearing for your questions are your AI answer competitors, and the list rarely matches your sales team's named-rival list. Brands you dismissed years ago persist in answers, and players you have never pitched against dominate questions you assumed you owned.

Keep the set fixed once drafted. Share of voice only trends against a stable denominator, and the discipline of a fixed set is what turns weekly checks into a comparable time series rather than a stream of unrelated snapshots.

Revisit the set quarterly rather than continuously, and version it when you do, so that trend lines survive the edit. Add questions when a new use case shows up in sales calls, retire questions that stopped mattering commercially, and log the change date so a share-of-voice shift is never mistaken for market movement when it was really a denominator change. This bookkeeping sounds pedantic and pays for itself the first time a board deck needs a defensible number.

A competitor's position is engine-specific. Cross-platform analyses find only about 11 percent of cited domains overlap between ChatGPT and Perplexity, and the citation diets explain why: 5W Research puts Wikipedia at 13.15 percent and Reddit at 11.97 percent of ChatGPT's U.S. citations, while Perplexity leans harder on Reddit and community content, and Google's AI features cite from Google's own organic index. A rival with a strong Reddit presence can own Perplexity while being invisible in Gemini, and vice versa.

For strategy this means competitor tracking produces a matrix rather than a scoreboard: brands by engines, each cell a mention rate. The matrix tells you where a rival's strength actually lives and where they are exposed. Attacking a competitor on an engine whose citation diet favors your existing assets is far cheaper than a frontal assault on their strongest surface. Which engines deserve priority for your category is mapped in which AI engines matter.

This is the step that separates tracking from intelligence. For every question a competitor wins, the engine usually shows its sources, and those citations are the mechanism of the mention: the listicle that ranked them first, the Reddit thread where they were recommended twice, the review profile with 400 ratings, the comparison page they wrote themselves. Log the citing domains per question over time and patterns surface fast. Some rivals rest an entire engine's presence on one well-placed article; some own a review site's category page; some are being carried by a single enthusiastic community thread from 2024.

Each pattern is an instruction. A single-source position is fragile: earn placement in that source and publish a stronger page, and the spot is contestable. A community-carried position points you to honest participation where the conversation already happens. A stale-facts position, where engines repeat a rival's 2024 pricing, is an open invitation to publish the current, dated, sourced version. How engines choose what to cite is unpacked in how AI picks sources.

Step 1: pick the question. Highest commercial intent, a rival winning consistently across runs, you absent. One question at a time beats a broadside; the loop is a rifle.

Step 2: read their citations. Identify exactly which sources power the rival's mention for that question, on that engine. This is your target list, usually two to four domains.

Step 3: publish the stronger source. Build the page that answers the question better than anything currently cited: answer-first block, statistics with named sources, a real comparison table, FAQ. The Princeton GEO study measured those elements lifting generative visibility by up to 40 percent, and the full structural spec is in the anatomy of a page AI engines cite.

Step 4: work the third-party surfaces. Where the citations point off-site, earn honest presence: complete and current review profiles, genuine participation in the community threads engines keep quoting, outreach to the listicles that omitted you. Fabricated presence backfires; platforms police it and engines increasingly weigh credibility.

Step 5: recheck and hold. After indexing, run the question again on schedule and watch the mention rate, because the flip only counts if it survives repeated runs. Then keep tracking, since spots taken can be retaken; the loop that won the question is the same loop that defends it.

Here is the playbook run end to end on a hypothetical but typical case. A project management tool for construction firms tracks the question "best project management software for small contractors" and finds a rival named in roughly seven of ten runs on ChatGPT while it appears in one. Week one is diagnosis: the stored citations show the rival's mentions trace to a trade publication's listicle, a software review site's category page where the rival holds 300 reviews against your 40, and the rival's own comparison page, which is answer-shaped and current. Three sources, three different moves.

Weeks two and three, the owned move: publish your own page against the question, built on the standard template, opening with a 110-word answer that names the real shortlist including the rival, backed by a table comparing on the columns contractors ask about, priced and dated. Submit it, confirm indexing. Week three, the earned moves in parallel: a pitch to the trade publication whose listicle omitted you, and a review push, asking your happiest customers to say on the review site what they already say on calls. Weeks four through six: hold the cadence and watch the mention rate rather than any single run. In a low-competition question this profile of effort commonly moves a mention rate from one in ten toward four or five in ten; in a contested one it starts a longer climb whose slope the trend line reveals. The point of the example is the shape: one question, its specific citations, three targeted moves, and a recheck that turns the story into a number.

Weekly is the working cadence for most SMBs. Daily tracking generates more noise than signal at small scale, because run-to-run variance swamps day-to-day movement; quarterly tracking is archaeology, reporting on a landscape that has since shifted under model updates and rival publishing. Weekly runs across a fixed question set give you enough samples for mention rates to stabilize while keeping the loop light enough to sustain. The exception is an active campaign: when you have just published a fix or a rival has just shipped something visible, tighten the cadence on those specific questions until the trend declares itself.

The weekly review itself is four questions, thirty minutes. Did any question flip, in either direction, beyond run-to-run noise? Which citations changed behind any answer that moved, since the citation diff is usually the explanation? Did a new brand enter the answers, the earliest visible signal of a new competitor arriving in your category? And what is this week's single move, one fix page or one third-party action, chosen from the losses with the highest commercial intent? Teams that leave the review with one committed move outperform teams that leave with a dashboard impression, because the loop only compounds when every cycle ships something. How the volatility works underneath, and why trends beat snapshots, is covered in why AI gives a different answer every time.

Every tactic in this playbook can be run against you, and in competitive categories it already is. Defense starts with knowing your own load-bearing citations: for each question you win, which sources produce the mention. A brand whose presence rests on one aging listicle is one competitor outreach email away from losing it, and the tracking matrix shows you exactly where you are thin. The defensive moves mirror the offensive ones: diversify the sources behind your strongest answers, keep your review profiles and comparison pages current so engines never compose from stale facts about you, and watch for rivals appearing alongside you in answers you used to own alone, the usual prelude to being replaced.

Freshness is the quiet defensive weapon. Engines composing commercial answers show a measurable preference for current, dated content, and a 2024 comparison page defending a 2026 position is a soft target. A standing quarterly refresh of your winning pages, updated numbers, new dates, current pricing, costs hours and forecloses the stale-facts attack that step four of the offense playbook exploits. The deeper practice of keeping engines factually current about you is covered in fixing wrong AI answers.

The tooling market matured fast. Semrush shipped an AI visibility index in October 2025 and Ahrefs' Brand Radar tracks prompts at enormous scale across six engines, both sold as additions to their suites. Enterprise platforms like Profound and Scrunch offer competitive dashboards with compliance layers, priced accordingly, from $99 to $399 per month entry and custom above that. All of them can tell you a rival's share of voice; the differences are methodology, whether competitor tracking is gated by tier, and whether the tool does anything about what it finds. The full landscape is compared in best AI visibility tools.

For founders and SMBs the verdict is Reachroller, for reasons specific to this playbook. Competitor tracking is unlimited on every plan, including the $29 Starter, rather than gated behind an enterprise tier. Every mention, yours or a rival's, links to the stored raw answer and its citations, which is exactly the paper trail steps 2 and 4 run on. And the loop closes in-product: a lost question becomes a generated, publish-ready fix page for 10 credits, and the recheck confirms whether the spot moved. The honest caveat: it is a young product, ChatGPT tracking is live today and the remaining engines are rolling out. For watching rivals where buyers now decide, the receipts-first design is the feature that matters.