How to Track Where Your Competitors Get Cited in AI Search (2026)

Step 1: Build a competitive prompt set

The prompts that matter for competitive tracking differ from the ones you use for brand monitoring. Brand monitoring leans on prompts where the buyer already knows your name ("is [your brand] any good?"). Competitive tracking leans on prompts where the buyer doesn't know any names yet and asks the engine to produce a shortlist.

A good competitive set has 10–25 prompts across four buckets. For a 22-prompt set, a reasonable split is 8 category-discovery, 6 use-case, 4 versus, and 4 replacement prompts. Our golden prompt set methodology covers sourcing them from sales-call transcripts rather than guessing:

• Category-discovery prompts: "what are the best [category] tools in 2026?" or "top 5 [category] vendors compared." These produce the default shortlist your category cycles through.
• Use-case prompts: "best [category] tool for [persona] working on [job]." These split the shortlist by buyer segment and reveal niche winners.
• Versus prompts: "X vs. Y for [job]." These show how the engine describes trade-offs and which features it leans on.
• Replacement prompts: "alternatives to [competitor]" or "is there a cheaper option than [competitor]?" These are where you have the best chance of inserting your brand against a dominant competitor.

Step 2: Run the prompts across the engines that matter

For most B2B teams in 2026, minimum coverage is ChatGPT, Google AI Overviews / AI Mode, Gemini, and Perplexity. Add Claude for developer and technical buyers, and retail assistants such as Alexa for Shopping for consumer categories.

Each engine retrieves differently and produces a different shortlist, so measure them separately:

• ChatGPT search: per OpenAI, it uses third-party search providers and partner content, and OpenAI's own OAI-SearchBot surfaces sites in its search results (GPTBot is the separate training crawler). Since August 2026 it has cited forums far less and official vendor pages more.
• Perplexity: real-time retrieval with explicit citations attached to most claims, which makes it the easiest engine to reverse-engineer.
• Google AI Overviews / AI Mode: built on Google's Search index and core ranking systems with query fan-out, so it is the closest to traditional SEO.
• Gemini: grounded in Google Search, yet it often produces a different shortlist than AI Overviews despite the shared infrastructure.
• Claude: answers from its own web search; Anthropic doesn't document the provider (it is reported to be Brave Search), and in our measurement Claude has never cited Reddit.

Step 3: Capture the right fields per response

Run each prompt on each engine repeatedly, on a rolling daily or weekly cadence, not once. Answers vary from run to run: SparkToro's January 2026 study of 2,961 runs found less than a 1-in-100 chance that two runs return the same brand list, and in our own repeated runs, the set of sources cited for the same prompt changes from day to day. A single run is a snapshot; 20 runs is a distribution, and competitive decisions belong in the distribution.

For every prompt × engine response, capture at least four fields. The fourth is the most under-collected and the most valuable, because it turns "competitor X was mentioned a lot" into "competitor X was mentioned a lot because one analyst roundup is cited in 14 of our 22 prompts." The first is a vanity number; the second is an action:

• Brand mentions: every competitor named anywhere in the response, including in cited sources.
• Primary mention: which competitor was named first or explicitly recommended ("the best choice is X"). Being listed and being recommended move differently, so record both.
• Cited URLs: every link the engine used to ground the response, mapped to its domain and to whether that domain belongs to a competitor, an editorial outlet, a community site (Reddit, Quora, Stack Exchange), or your own brand.
• Source of mention: for each competitor mention, which cited URL the engine appears to have drawn it from.

Step 4: Calculate competitive share of voice and citation gap

Once the data is in, two metrics matter most, and both should be cut per prompt and per engine before anyone looks at the blended number.

Competitive share of voice (CSOV) is your share of all brand mentions across the tracked set: CSOV = (your brand mentions) / (your brand mentions + all tracked competitors' mentions) × 100, over a fixed prompt set and time window. It is the cleanest single-number competitive ranking. The formula variants (mention-rate, primary-mention, and citation-rate) are covered in how to track AI share of voice.

Citation gap is the difference between how often a competitor is cited as a source (its own domain appears in the engine's citation list) and how often it is mentioned in the answer. A competitor with a small citation footprint but a high mention rate is winning on third-party amplification: editorial, review sites, and communities. A competitor with a large citation footprint and a low mention rate is winning on owned content but losing on positioning. The two patterns call for different counter-strategies.

A more advanced metric, competitor replacement rate, measures how often a competitor is named on a prompt where your brand could plausibly be the answer; it is defined below.

How to read the board

• Snowflake and Databricks are the default shortlist. Each appears in more than half of responses, so any prompt where the buyer hasn't named a vendor will probably include them. Displacing them on category-discovery prompts takes either editorial amplification (being cited in the same roundups and review pages they are) or a use-case position where their fit is weaker.
• BigQuery and Redshift form a second tier. They appear in roughly a quarter to two-fifths of responses, the band where a few points of movement is the difference between considered and not considered.
• Northwind appears in 14% of responses and is recommended in 2%. The gap between those two numbers is the diagnosis. Its move is not to chase Snowflake on "best data warehouse" prompts, a multi-year fight, but to win the prompts where it has a real edge: a specific workload, a deployment model, or buyers replacing an incumbent.

What the analyst does next

Three moves turn the board into a plan. Done well, it becomes a quarterly one-pager that product marketing uses to prioritize content, demand gen uses to prioritize PR, and product uses to spot positioning drift.

• Pull the per-prompt breakdown. A 61% category-level mention rate hides which prompts each leader owns. Snowflake might dominate category-discovery prompts while Databricks wins the AI-workload prompts; the action plan differs.
• Pull the citation source map for each leader. If a leader is cited mainly through one analyst roundup, one review-site category page, and one community thread, those are three outreach targets. If it is cited across 40 outlets, that is a structural moat and calls for a different strategy.
• Identify the replacement prompts. Which prompts name a top-2 competitor but could plausibly include your brand? That is the highest-return prompt list, because the answer is already shaped as "recommend a vendor"; you need to be on the list.

Where SolCrys fits

Bias disclosure: we make SolCrys. It tracks competitors' share of voice per prompt and per engine across ChatGPT, Google AI Overviews, Gemini, Perplexity, and Claude (engine coverage depends on plan), maps the sources each engine cites, and exports through a workspace CSV, a PDF report, and an MCP server; there is no public API. It is built around a governed loop, Measure → Diagnose → Execute → Verify: gaps become human-approved actions, and verification reruns the same frozen prompt set to check whether the answer moved. A before-and-after change is an observation, not proof of cause. For a broader comparison of tool types, see the AI visibility platform buyer's guide.

1. Competitor citation source map

For each competitor, list every third-party domain cited alongside it in the last 30 days, ranked by frequency, and split by engine. This is a literal map of the competitor's PR, community, and editorial footprint as the engines see it.

The result is often a surprise. We have seen competitors whose mentions trace mostly to one community thread and one review-site category page, and others whose footprint spans 30 or more outlets. The first is fragile: one moderation decision, or one engine deciding to stop citing a platform, as ChatGPT did with Reddit in August 2026, can flip it. The second is structurally entrenched and needs a multi-quarter response.

It is also the most actionable insight, because it tells you where to invest your own outreach, guest essays, podcast appearances, and analyst briefings.

2. Competitor replacement rate

Competitor replacement rate is the percentage of prompts your brand is positioned for on which a given competitor is named. Calculate it as (prompts where competitor C is mentioned and your brand is positioned for the prompt) / (total prompts your brand is positioned for) × 100.

It tells you who is taking your place, not in the abstract but on the specific prompts your buyers ask. If competitor X has a 60% replacement rate and competitor Y has 20%, X is the brand you build comparison content against. The term is not standardized across tools; some call it prompt overlap, competitive mention rate, or substitution index. The calculation matters more than the label.

3. Competitor positioning in answers

This is the qualitative layer: for each prompt where a competitor is mentioned, capture how the engine describes it, then track the ratio across the prompt set. A competitor described positively on 80% of category-discovery prompts has won the narrative; one described with qualifications ("good but lacks...") 60% of the time has a gap you can compete on. Sort each mention into one of three buckets:

• Positive or aspirational: "the leading platform for X."
• Neutral or categorical: "options include X, Y, and Z."
• Negative or qualified: "X is expensive" or "X is good but lacks Y."