Competitor AI Visibility: How to Benchmark Share of Answer in 6 Steps (and Find the Shortlists Your Brand Loses)

Competitor AI Visibility: The Benchmark That Shows Where the Brand Loses

Pew Research Center tracked 68,879 Google searches in March 2025 and found that users clicked a traditional result in only 8% of visits when an AI summary appeared, versus 15% without one — and clicked a link inside the summary just 1% of the time (Pew Research Center). The answer text, not the click, now carries the shortlist. Competitor AI visibility is the discipline of measuring who occupies that answer text, and this section establishes the benchmark that exposes where a brand loses its place.

Why a Single-Brand Score Cannot Be Invoiced

Most AI visibility tools return one number for one brand. An agency cannot invoice a score; it can invoice a displacement chart showing the client at 18% share of answer against a rival at 34%. That contrast is the deliverable.

What the Benchmark Measures

The outcome is precise: a competitor benchmark that shows where the brand loses shortlist positions across ChatGPT, Perplexity, Gemini, Copilot, and Google AI Overviews, and which cited sources are handing those positions to rivals. Alef's AI visibility workspace tracks mentions, citations, rankings, sentiment, and competitor share across the same answer engines, so the six-step procedure below mirrors what the platform runs continuously rather than a manual exercise.

The first pass on one client takes roughly two to three hours, requires no engineering work, and produces a repeatable artifact that can be re-run monthly. Where the diagnostic on brand invisibility explains why a brand disappears from AI answers, this benchmark quantifies who took its place.

When a Competitor AI Visibility Benchmark Is the Right Deliverable

Four situations make a competitor benchmark the correct deliverable rather than a generic visibility report.

• A forwarded screenshot. A client sends a ChatGPT or Perplexity answer that names three rivals and omits the brand. The request that follows is "prove where we stand," not "write more blogs."
• A renewal or QBR. The agency needs an AI line item it can defend. A single-brand score invites the question "compared to whom?"; a competitor-anchored benchmark answers it before it is asked.
• Sliding organic click-through. As AI summaries absorb queries, Google users click a link in 8% of visits with an AI summary versus 15% without one, per Pew Research Center. Clients want to know whether that loss is going to a named rival.
• AI-referred traffic already in analytics. Once the channel is live and measurable — see how to isolate it in AI-referred traffic and how to measure it — a benchmark has a baseline to trend against.

Prerequisites and the Premature Case

The benchmark requires an indexable site with crawlable content, a defined category, at least three named competitors, and a shared prompt set of 20 to 50 buyer questions. No code is involved, but spreadsheet fluency and prompt design are assumed. A pre-launch site with no indexable content is premature: answer engines cannot retrieve or cite what they cannot crawl, so fix crawlability and indexation first. Agencies running several clients should pair the benchmark with the multi-client AI visibility operating model so each benchmark feeds a repeatable workflow.

The 6 Steps, Broken Into 12 Executable Actions

The six steps below are not a one-off audit. They describe a repeatable procedure an agency can run for every client in its roster, on the same cadence, so that results are comparable across accounts and across months. The outline names six steps; in practice, several steps split into two executable actions each, which yields twelve discrete actions. Each action has a defined input, a defined output, and a verification check so the analyst knows the action is complete before moving on.

A note on effort before starting: a full first pass across five engines and 30 prompts takes roughly 12 to 16 analyst hours for a single client, most of it in prompt design and source logging. Subsequent monthly runs compress to 4 to 6 hours because the prompt set is frozen and only the answer capture and scoring repeat.

Step 1 — Select the competitor set

1. Define and freeze a 3-to-5 brand competitor set, split into direct and adjacent rivals. The competitor set is the foundation of the entire benchmark; if it is wrong, every downstream number is wrong. The set should contain three to five brands, no more, because each additional brand multiplies the manual scoring work in Step 4 without adding proportional insight.

Split the set into two tiers:

• Direct competitors — same category, same buyer, same purchase occasion. If the client sells project management software to mid-market teams, a direct competitor is another project management platform sold to the same buyer.
• Adjacent competitors — different category, same underlying problem. The same client's adjacent competitor might be a spreadsheet suite, a whiteboard tool, or a professional services firm that solves the same coordination problem through a different product category.

Adjacent competitors matter because AI answer engines routinely mix categories in a single response. A prompt asking for "the best way to coordinate a 40-person product team" may surface a project management platform, a spreadsheet template, and a consultancy in the same answer. If the benchmark only tracks direct rivals, it will miss the displacement happening from outside the category.

Freeze the list in a dated document. Record the date, the analyst's name, and a one-line rationale for each brand's inclusion. The freeze matters because competitor sets drift; a brand added in month three invalidates month-one and month-two comparisons unless the entire history is re-scored. A frozen list with a version number (for example, "Competitor Set v1, frozen 2025-01-15") lets the agency add a brand later as v2 and re-score only the prompts where the new brand appears.

Exclude marketplaces and review aggregators from the brand list. G2, Capterra, Amazon, and similar platforms will appear constantly in AI answers, but they are not competitors in the displacement sense; they are sources. They belong in the source analysis in Step 5, not in the brand scoring in Step 4. Mixing them into the brand set inflates the denominator and makes the client's share of answer look artificially small.

Step 2 — Build the shared prompt set

2. Draft 20 to 50 buyer questions tagged by intent, written the way a buyer actually types them. The prompt set is the measuring instrument. A benchmark built on 10 prompts is a sample of ten; a benchmark built on 50 is a sample of fifty. The practical range for an agency engagement is 20 to 50 prompts, with 30 as a workable default for a mid-market client and 50 for a category with many sub-segments.

Write prompts the way a buyer types them, not the way a marketer would phrase them. "Best enterprise project management software for regulated industries" is a marketer's prompt; "what project management tool do banks use" is closer to a buyer's. The prompt set should include at least three phrasings per intent cluster:

• "Best X for Y" — discovery and shortlist formation, for example "best project management software for healthcare teams."
• "X vs Y" — comparison, for example "Asana vs Monday for a 50-person team."
• "Is X worth it" — buying validation, for example "is Notion worth it for a 200-person company."

Tag every prompt by intent: discovery, comparison, or buying. The tag drives the intent weighting in Step 8, so it must be assigned at prompt creation, not retrofitted after the answers come back. A prompt that does not fit cleanly into one of the three tags is usually two prompts in disguise and should be split.

The prompt set runs identically for every brand in the frozen competitor set. This is the single most important discipline in the benchmark: the same prompt, run at the same time, on the same engine, scored against every brand. If the client's prompts differ from the competitors' prompts, the resulting share-of-answer figures are not comparable and the benchmark is invalid.

Step 3 — Run the prompt set across engines

3. Execute the identical prompts in ChatGPT, Perplexity, Gemini, Copilot, and Google AI Overviews, logging the run conditions. Five engines is the practical minimum for a credible cross-engine view. Each engine draws on a different index, a different retrieval layer, and a different ranking logic, so a brand that wins on one engine can be invisible on another.

For every run, log:

• Date and time of the run, in a single timezone.
• Engine and model version — for example, "ChatGPT, GPT-4o" or "Perplexity, Sonar." Model versions change and answers drift with them.
• Account state — logged-in versus logged-out, personalization enabled or disabled, region. A logged-in account with a browsing history will produce different answers than a clean session, and the difference is not random noise; it is a different retrieval context.
• Prompt ID from the frozen set, so the answer maps back to its intent tag.

The reason for this logging discipline is that answers drift between runs. A single snapshot is a sample of one, and a sample of one cannot support a claim like "the client is losing to Competitor B on Perplexity." Running each prompt three times per engine and recording all three answers gives a stability read: if the client appears in two of three runs, that is a different finding from appearing in zero of three. For agencies running this at scale, Alef's prompt intelligence workspace automates the capture and versioning so the analyst is not manually screenshotting five engines per prompt.

The expected outcome of Step 3 is a raw answer log: every prompt, every engine, every run, with the full answer text captured verbatim. Verification: the log should contain prompts × engines × runs rows, with no blank cells. For 30 prompts, five engines, and three runs, that is 450 captured answers.

Step 4 — Record mention and citation separately for each brand

4. Score every answer for brand mention and brand citation as two independent variables. This is where most single-brand AI visibility tools collapse the data and lose the insight. A mention and a citation are not the same thing, and treating them as one number hides the displacement pattern the client is paying to see.

• A mention is the brand name appearing inside the answer text, whether or not any link is attached.
• A citation is an attributed link to a source that the engine used to construct the answer.

The two diverge sharply. Ahrefs' analysis of 75,000 brands found that only about 28% of AI brand mentions include a link, with Perplexity at roughly 51.6% and Google AI Overviews at about 10.7% (Ahrefs AEO course, lesson 1.3). In other words, the majority of the time a brand is named in an AI answer, no clickable path back to the brand exists. For an agency, that means a mention-only score overstates the traffic opportunity, and a citation-only score understates the brand's presence in the answer.

Score each answer on a four-cell grid per brand:

The "mentioned but not cited" cell is the most common and the most under-reported. The "cited but not mentioned" cell is the ghost citation, and it is the subject of Step 6. The scoring output is a per-answer, per-brand record of which cell applies, which becomes the input to Step 5.

Step 5 — Compute share of answer

5. Calculate each brand's share of answer as mentions divided by total brand mentions, reported unweighted and position-weighted side by side. Share of answer is the headline metric of the benchmark. It answers the question the client actually asks: "When a buyer asks an AI engine about our category, how often does our brand appear relative to everyone else?"

The calculation is straightforward once the scoring in Step 4 is complete:

1. Count the total brand mentions across the frozen prompt set. If a single answer names three brands, that is three mentions.
2. Count each brand's mentions within the same answers.
3. Divide each brand's mentions by the total to get unweighted share of answer.

Report the position-weighted figure alongside the unweighted one. Position weighting reflects that a brand named first in an answer carries more shortlist weight than a brand named fifth in a list of five. A simple weighting scheme assigns a score of 1.0 to the first-named brand, 0.7 to the second, 0.5 to the third, and 0.3 to any brand named fourth or later, then normalizes. The exact weights are a judgment call; the discipline is to apply the same weights to every brand and to report both numbers so the client can see whether the client's position is strong-but-rare or weak-but-frequent.

The critical presentation rule: show the client's number next to each rival's, never in isolation. A share of answer of 12% means nothing until the client sees that Competitor A holds 31% and Competitor B holds 24%. The displacement view is the deliverable. For a deeper treatment of the calculation methodology and its edge cases, the guide to calculating share of voice in AI search correctly covers the normalization choices in detail.

Step 6 — Identify the sources competitors are cited from

6. Log every cited domain per answer, count frequency, and separate owned domains from earned ones. This step converts the benchmark from a scoreboard into an action plan. The question shifts from "who is winning" to "where are the winners being cited from, and why is the client absent from those sources?"

For every answer in the log, record each cited domain. Then aggregate:

• Frequency per domain across the full prompt set and per engine.
• Owned versus earned — is the domain the brand's own site, or a third-party publisher, review platform, forum, or directory?
• Brand attribution — does the cited page name the brand, or is it a ghost citation?

The ghost-citation pattern is larger than most agencies expect. Semrush's ghost-citations study found that 62% of AI citations carry no brand mention at all (Semrush Ghost Citations Study). A cited-but-unnamed source is still a winnable slot: if the engine is reaching for a third-party comparison page that names three competitors and not the client, that page is an outreach target, and getting the client added to it moves the client into the answer without requiring the client to outrank anyone on its own domain.

The output of Step 6 is a source gap table: for each high-frequency cited domain, which brands appear on it, and whether the client appears. That table is the direct input to Step 7. The scale of the source layer is worth noting: Search Engine Journal's analysis of 24 million requests found ChatGPT now crawls 3.6 times more than Googlebot (Search Engine Journal), which means the set of pages AI engines can reach is expanding faster than most brands' content production, and the source gap is a moving target rather than a fixed list.

Step 7 — Convert the source gap into a task

7. Assign a content task or an outreach task for every source where rivals appear and the client does not, with an owner and a due date. A benchmark that ends with a chart is a report; a benchmark that ends with a task list is a program. Step 7 is where the share-of-answer finding becomes billable work.

For each row in the source gap table, assign one of two task types:

• Content task — the engine keeps reaching for a page type the client does not have. If competitors are cited from "best X for Y" listicles on their own domains and the client has no equivalent page, the task is to publish that page. The task description names the prompt cluster it targets, the page type, and the internal linking requirement.
• Outreach task — the engine is citing a third-party domain the client cannot control. If a review platform, industry publication, or comparison site is cited repeatedly and names three competitors but not the client, the task is to earn a placement on that domain. The task description names the domain, the specific page, and the placement type (listing, review, contributed article, directory entry).

Every task carries an owner and a due date. Tasks without an owner are wishes. The due date should be realistic against the content production calendar; a content task that requires original research gets four to six weeks, an outreach task that requires a publisher relationship gets six to eight weeks.

The expected outcome of Step 7 is a task board with one row per source gap, each row tagged content or outreach, each row owned and dated. Verification: every high-frequency cited domain where a rival appears and the client does not should have exactly one open task against it. If a domain has no task, the analyst has decided not to pursue it, and that decision should be recorded with a reason.

Step 8 — Weight by intent

8. Publish both the raw share of answer and the buying-intent share so the client sees which number moves revenue. The final action reframes the benchmark around commercial outcome. A mention in a discovery prompt ("what is project management software") builds awareness; a mention in a buying-intent prompt ("best project management software for a 50-person healthcare team") sits closer to a purchase decision. The two are not worth the same, and reporting them as a single blended number hides the distinction that justifies the engagement.

Recompute share of answer using only the prompts tagged buying in Step 2. Report the buying-intent share alongside the raw share for every brand in the set. The comparison is often revealing: a client may hold a respectable 18% raw share but only 6% buying-intent share, which means the brand is visible in the category conversation but absent at the moment of shortlist formation. That gap is the strongest possible argument for the content and outreach work in Step 7.

The intent-weighted view also protects against a common failure mode: a client that dominates discovery prompts through broad brand awareness but loses every comparison and buying prompt to a smaller, more focused rival. Without the intent split, the blended number flatters the client; with it, the displacement is visible and the remediation is specific.

The output of Step 8 is the final benchmark summary: a table of brands, raw share of answer, buying-intent share of answer, and the delta between them. That table is the artifact the agency presents, and it is the artifact that makes the case for the next quarter's work.

The benchmark is complete when every brand in the frozen set has a raw share, a buying-intent share, a source gap table, and an owned task list. Anything short of that is a diagnostic, not a benchmark.

Common Mistakes That Invalidate a Competitor Benchmark

Most flawed benchmarks fail on method, not effort. Each of the following errors is recoverable if caught before the first report ships.

The errors that matter most

• Freezing the competitor list too late. Adding a rival mid-cycle changes the denominator of share of answer and makes every prior month incomparable, so the set must be locked before the first run.
• Scoring mentions and citations as one number. Ahrefs reports a 28% link rate for AI citations, meaning the two diverge sharply by engine; a blended score hides exactly where the client is losing (Ahrefs).
• Running each prompt once. Identical prompts return different answers across runs, so a single snapshot is an anecdote, not a measurement — the same reproducibility problem covered in rank tracking accuracy and why tools report positions you cannot see.
• Counting only owned domains in the source analysis. Semrush's 62% ghost-citation rate means the most winnable slots are third-party pages that never name a brand (Semrush); the mechanics are set out in AI citation tracking for mentions without links.
• Reporting the client's score without the rivals on the same axis. A standalone number invites the question "is that good?", which the benchmark exists to answer.
• Treating marketplaces and review sites as competitors. G2, Capterra, and Reddit threads are sources to win, not brands to displace, and mixing them corrupts the share calculation.
• Skipping the intent weight. A discovery mention reported as equal to a buying-intent mention inflates the client's position and erodes trust at renewal.

Verification checklist

Before publishing, confirm that the competitor set is versioned and dated, mentions and citations are stored as separate fields, every prompt has at least three logged runs, the source table includes non-owned domains, and every client figure sits beside its rivals.

Benchmark Summary: Steps, Inputs, and Outputs

The table below condenses the full twelve-action procedure into a single reference an agency team can run without reopening the earlier sections. Each row carries a concrete input, a named output artifact, and a verification check, so a reader who sees only this table can still execute the benchmark end to end.

The verification column matters more than it appears. A benchmark that cannot be re-run against the same frozen competitor set and prompt library produces numbers nobody can defend in a client review, which is the failure mode the full step-by-step benchmarking walkthrough is designed to prevent. The citation-share rows deserve particular scrutiny: Semrush's ghost citations study found that 62% of AI citations do not lead to a brand mention, so a domain can be cited heavily on a client's prompts while the brand itself never surfaces in the answer.

What the Benchmark Proves and What to Do Next

The six steps form a single operating loop rather than a one-off audit: freeze the competitor set, run the shared prompts, record mentions and citations as separate signals, compute share of answer, map the sources engines cite, and convert the gap into a content or outreach task. Each pass through the loop sharpens the next.

The commercial argument is straightforward. A benchmark turns an invisible problem into three fundable specifics — a named rival, a specific prompt, and a specific source. Clients approve work against those, not against a single-brand score. The ghost-citation gap makes the source layer especially urgent: Semrush's Ghost Citations study found 62% of AI citations occur without a brand mention, so a competitor can hold a shortlist position through sources the client never appears in.

Cadence matters as much as method. Monthly re-runs on a frozen competitor set and prompt list produce a trend line that survives client scrutiny; prompt changes are versioned and dated so any shift in share of answer can be attributed to content work rather than a moving baseline. For agencies that want the metric definitions behind each figure — mention rate, citation rate, share of answer — measuring AI visibility metrics that predict pipeline explains how each one maps to downstream demand.

Key takeaways - Freeze the competitor set and prompt list before the first run; version every change. - Record mentions and citations separately — they diverge, and the split is diagnostic. - The source gap, not the score, is the actionable output of the benchmark. - Report displacement: which shortlist positions the brand loses, to whom, and from which cited sources.

Frequently Asked Questions About Competitor AI Visibility

How many competitors should a competitor AI visibility benchmark track?

Three to five, split between direct rivals and adjacent players that appear in the same consideration set. Every addition dilutes the denominator of share of answer and multiplies the prompt runs, because each engine call must be repeated per brand per prompt. A benchmark with eight competitors and 40 prompts across five engines produces 1,600 runs per cycle; the same scope with four competitors produces 800. Agencies that track too wide end up with thin per-brand samples and cannot defend the percentages in a client review. The practical split is two or three direct competitors the client names in every pitch, plus one or two adjacent players that answer engines surface as substitutes.

How is share of answer different from share of voice in AI search?

Share of answer counts how often a brand appears inside a fixed prompt set, with each mention weighted equally. Share of voice applies position weighting and engine weighting on top, so a recommendation in the first sentence of a ChatGPT answer outranks a footnote mention in a Copilot response. Both metrics should be reported side by side: share of answer is the stable, auditable number clients can verify, while share of voice captures the quality of the placement. Reporting only one hides the difference between being listed and being recommended, which is the distinction that determines whether the client enters the shortlist.

Why do competitors get cited when the brand does not?

Answer engines retrieve from sources they already associate with the topic, and the cited page — not the brand — is the unit being selected. Semrush's ghost citations research found that 62% of AI citations carry no brand mention at all, which means a competitor can win a shortlist position through a third-party comparison page, review roundup, or documentation article that never names them in the visible answer. The brand's own domain may rank well in classic search and still lose the citation because no trusted page connects the entity to the query. This is why source-level analysis, covered in the guide to AI visibility tracking and measurement, matters more than the aggregate score.

How often should competitor AI visibility tracking be re-run?

Monthly, on a frozen prompt set, with any prompt change versioned and dated. Answer engines update models and retrieval indexes on their own schedules, so an unversioned prompt set produces month-over-month movement that cannot be attributed to content work or to model drift. Freezing the set means the client sees a clean trend line; versioning means the agency can explain a step change when a prompt is added or reworded. Quarterly deep runs that expand the prompt set can sit alongside the monthly baseline without contaminating it.

Can competitor AI visibility be tracked without a paid tool?

Yes, manually, at roughly two to three hours per client per cycle for a four-brand, 40-prompt benchmark across five engines. The cost is consistency, not capability: manual runs drift as the prompt set grows, because the person recording mentions interprets "mentioned" differently each cycle, and screenshots do not aggregate into a share-of-answer figure without a spreadsheet built for it. Manual tracking is defensible for a single client during a pilot. It stops being defensible at the point where the agency needs to compare ten clients on the same metric definition.

Which engines must be covered?

At minimum ChatGPT, Perplexity, Gemini, Copilot, and Google AI Overviews, because citation behavior differs sharply by engine. Ahrefs' analysis of AI visibility types reports that Perplexity links to sources in roughly 51.6% of answers versus 10.7% for AI Overviews, which means a brand can be visible in one engine's answer text and absent from another's citation layer entirely. Tracking only ChatGPT produces a benchmark that misrepresents the client's position in Google surfaces, where AI Overviews now sit above the organic results. The guide to tracking brand mentions in ChatGPT and Perplexity covers the per-engine recording differences in detail.

Run the First Benchmark With Alef

The benchmark is the argument; the workspace is the instrument. Setup requires three inputs — the client's domain, three to five named competitors, and the shared prompt set — after which the displacement view reads across ChatGPT, Perplexity, Gemini, Copilot, and Google AI Overviews, with mentions, cited sources, and the resulting next actions held in one place. The AI visibility platform built for competitor-anchored benchmarking is where that first run begins. One client, one prompt set, one benchmark: enough to show exactly which shortlist positions are being lost, and to whom.

Sources

• Pew Research Center — Google users are less likely to click on links when an AI summary appears in the results
• Ahrefs — What Is AI Visibility? The 3 Types Every Marketer Needs to Know
• Semrush — Why 62% of AI citations don't lead to brand mentions (Ghost Citations Study)
• Search Engine Journal — ChatGPT Now Crawls 3.6x More Than Googlebot: What 24M Requests Reveal
• TechCrunch — Sam Altman says ChatGPT has hit 800M weekly active users
• Semrush — AI Overviews Study: What 2025 SEO Data Tells Us About Google's Search Shift
• Search Engine Land — Google AI Overviews now show on 13% of searches: Study