A competitor showing up in one AI answer is a reason to take a closer look, and it doesn't tell you that you've lost the topic. This guide is for marketing leads who already have tracked prompts and a competitor set, and who need to decide which losses deserve real work. By the end you'll have a short, ranked backlog where every item has a prompt, an engine, the observed counts, the answer and source evidence behind it, a proposed fix, an owner, and a way to retest.
To follow along you'll want the tracked prompt results, the full answers and cited sources behind them, the collection dates and engines, and your configured competitors. Reading prompt level win loss data well comes down to comparing like with like, keeping a few terms apart, and checking the answers before you trust the averages. The thresholds and labels below are working practices I'm suggesting, not published benchmarks, because nobody has agreed on a standard win/loss metric for AI answers.
Lock the prompt, engine, and comparison window
Compare like with like before you decide a competitor is winning. For each analysis, write down the exact prompt text, the AI engine (and the mode, if you know it), the date range, the language and location, and whether the prompt names a brand. Also note how many answers were valid, how many collection attempts failed, and whether the prompt or setup changed during the period.
Prompt analytics are easiest to read over a fixed window. Start with a 30-day view for the overall pattern, look at the most recent 7 days for movement, and use 90 days for context if the setup stayed comparable. These windows aren't a guarantee of statistical reliability.
Keep each engine's results visible, since a strong result on one can hide a loss on another. Split branded prompts from unbranded ones too, since an answer to a prompt that names your brand will often just repeat it, which isn't an unprompted endorsement.
You're done when any row can be understood without guessing which question, engine, period, or setup it came from. Two periods should use the same prompts, or you should say which were added or dropped.
The usual mistake is comparing your own manual ChatGPT session against an automated collection. ChatGPT Search can rewrite your question, use saved memories, and use your location, so a manual spot check isn't necessarily what the tracker saw. If you run the same prompt across several engines, track the same prompts across ChatGPT, Gemini and Perplexity with one fixed setup so the comparison stays fair.
Open the answers behind the prompt metrics
Use the prompt row to find the pattern, then read the actual answers to check what happened. Start with prompts where your mention rate or citation rate is low and a competitor shows up often. For each one, read more than one answer, spread across different dates and the engine you care about.
For every answer, capture whether your brand and each competitor are named, and whether yours is recommended, only listed, described, or left out. Mark each visible source link as yours, a competitor's, or an independent publisher's. Check brand names and aliases too. A company name that is also an ordinary word can produce false matches, so verify important rows against the answer text and don't quietly change the dashboard's definition.
This is the first of a few places where DeepSmith does the work. In DeepSmith, the Prompts area gives you per-prompt mention and citation rates along with the full answer history. So the way to work is to open the prompt, look at the platform breakdown, and then read the answers behind the apparent loss. The answer view shows where your brand was named, how it was described, and which sources the answer cited.
You're done when each important number has answer evidence behind it, and you've read at least the answers that drive your diagnosis. With a small result set, read all of them. With a large one, read a spread of dates and outcomes, because picking only the worst screenshot gives you a distorted view.
Common mistake: "The competitor appears and we don't" describes one answer you observed. It doesn't prove that the competitor took a slot that would otherwise have been yours. Other common errors are treating a mention as a recommendation, counting an independent article as a citation of your own page, and trusting an automatically extracted label on an ambiguous answer without reading it.
Classify presence and citation outcomes separately
Label what happened in each answer before you average anything. For each pair of your brand and one competitor, give the answer one of four presence labels.
| Your brand named? | Competitor named? | Label | How to read it |
|---|---|---|---|
| Yes | No | Brand-only | You appear and this competitor doesn't |
| No | Yes | Competitor-only | This competitor appears and you don't |
| Yes | Yes | Both | Read the recommendation, the order, and any qualifications |
| No | No | Neither | Check whether other brands filled the answer, or no brands were named |
Then do the same for citations, using links to each brand's own pages, and keep the two tables apart instead of folding them into one score. A response can name your brand without linking to your site, and it can cite your page without naming you in the text.
Add a recommendation field as well, since "both" is not automatically a tie. Useful labels are recommended for the use case, included as an alternative, mentioned without a recommendation, explicitly qualified or discouraged, and unclear. These are worksheet suggestions, not DeepSmith categories.
An independent review page might recommend your brand while the citation belongs to the review site. Record that as a mention plus a third-party source, and don't count it as a citation of your own page. The piece on AI citations vs brand mentions goes deeper on the difference.
In any prompt level win loss review, failed or unavailable collection results belong outside this table. An answer that was captured and doesn't mention your brand is a real absence, and a timeout is not. If you also track Google AI Overviews, separate "no overview appeared" from "an overview appeared without our brand."
You're done when every valid answer has a presence label, a separate citation label, and enough context to judge the recommendation. The common slip is to call "both" a loss because the competitor was listed first, or to call "neither" a competitor win.
Calculate rates with visible denominators
Show counts next to percentages so the size of the gap stays honest. Let N be the number of valid, completed answers for one prompt, one engine, and one period. Your mention rate is the answers that mention you divided by N. The competitor-only loss rate is the answers that mention the competitor but not you divided by N, and the brand-only win rate, both rate, and neither rate work the same way. The owned-page citation rate is the answers citing at least one of your pages divided by N. The mention-rate gap is the competitor's mention rate minus yours, in percentage points.
Count each brand once per answer. If your name appears five times in one response, that's one presence, and several links to the same domain are still one citation success. When you reproduce a number from an AI visibility tool, match that tool's published definition, because vendors differ. Peec, for example, divides brand mentions by total brand mentions for share of voice, but divides responses mentioning the brand by total responses for visibility. Tools that track competitor share of voice in ChatGPT vary the same way. DeepSmith describes share of voice as visibility relative to competitors, and what's published doesn't spell out the counting formula, so don't label a homemade calculation as DeepSmith's share of voice. A head-to-head rate, brand-only answers divided by brand-only plus competitor-only answers, leaves out "both" and "neither" on purpose, so show that denominator.
A worked example
These numbers are made up for teaching. They're not DeepSmith customer results and they're not research findings. Say one prompt on one engine, over one window, produced 20 valid answers, and two more attempts failed and are recorded separately.
| Mention outcome | Answers | Share of valid answers |
|---|---|---|
| Your brand only | 3 | 15% |
| Competitor only | 11 | 55% |
| Both | 4 | 20% |
| Neither | 2 | 10% |
| Total | 20 | 100% |
Here's how to read it. You appear in 7 of 20 answers, a 35% mention rate. The competitor appears in 15 of 20, or 75%, so the mention-rate gap is 40 percentage points in its favor. The competitor-only loss rate is 11 of 20, or 55%. The 4 "both" answers need a human read, and the 2 "neither" answers aren't competitor wins. Your head-to-head rate is 3 of 14, about 21.4%, a different metric from your mention rate. The 2 failed attempts don't become absences, so collection completeness is 20 of 22 attempts, about 90.9%.
If your pages are cited in 3 of the 20 answers and competitor pages in 8, the citation rates are 15% and 40%, but check the answer-level overlap before you work out a competitor-only citation loss. If your mention rate was 25% last period and is 35% now, that's a 10 percentage point change, or 40% relative growth, so say which one you mean.
DeepSmith's analytics windows are 7, 30, and 90 days, and rates come back as fractions between 0 and 1, so 0.35 means 35%. You're done with this step when every percentage has a numerator and a denominator, the missing-data treatment is written down, and citation rate and citation share carry different labels.
Check how stable the result is before you trust it
A single answer is a sample, and these tools don't always give the same answer twice. SparkToro ran a study with 600 volunteers, 12 prompts, and 2,961 combined runs across ChatGPT, Claude, and Google AI, and found less than a 1 in 100 chance of getting an identical list of brands twice in the ChatGPT and Google AI comparisons it tested. That is about whole lists, and one hospital still appeared in 69 of 71 ChatGPT responses to a related question, so a single brand can stay steady while the list changes. The study is exploratory, so treat it as a caution. The research paper Don't Measure Once found that cited sources overlapped only partly from day to day, which supports looking at repeated observations. For more, see day-to-day noise in AI visibility scores.
There's no universal number of runs after which a win or loss becomes certain, so show the counts and call small or mixed samples provisional.
Trace repeated losses to the cited pages and answer claims
Treat the competitor's cited page as evidence of what the answer drew on, and don't treat it as proof of how the engine reasoned. For losses that keep coming back, put the answer, the competitor's page, any third-party sources, and your closest page side by side. Ask what exact buyer requirement the answer addresses, whether the competitor is recommended for it or only mentioned, which page is linked near the claim, and whether that page actually supports the statement. Then ask whether you have an accurate, findable page on the same requirement, or whether the problem is really missing evidence, a third-party source, an outdated description, or a product-fit difference.
| What you see | What it might mean | What to do next |
|---|---|---|
| Competitor named again and again, you're absent | A fit, entity, source, or coverage problem | Check fit and source evidence before commissioning content |
| You're named, but competitor pages get the citations | A citation gap, not total absence | Improve your best relevant page and see why other sources are stronger |
| Both named, but your description is wrong | A description or source-quality problem | Correct your authoritative page and find the sources repeating the error |
| Citations mostly go to review sites or directories | Reliance on third-party sources | Check listings and independent coverage, since a blog post alone may not close it |
| Your page can't be crawled or indexed properly | A technical eligibility problem | Fix access and indexing first |
| Competitor meets a requirement you don't | A real product-fit gap | Send it to product or positioning, and don't claim a capability you lack |
This is the second place DeepSmith helps. Its competitor citation views show the exact pages that earn citations and how each competitor performs by engine, tied to citation rate, share, trend, and the prompts that cite them. The Pages view does the same for your own pages. That moves you from "we lose this prompt" to "this page keeps getting cited for this requirement," though it won't show why a model made its choice. For one URL, page-level citation diagnostics describes what a tool should tell you, and auditing content for AI citations covers the page-side comparison.
You're done when your diagnosis has an evidence trail of prompt, answer excerpt, linked source, the buyer requirement, and the page (or missing evidence) on your side, and when you've said what's still uncertain. Don't assume a heading or word count caused the competitor's inclusion, because a report like this doesn't isolate cause.
Rank losses by business relevance, persistence, and fixability
Fix the losses that matter to the business and have a defensible fix, and don't simply go after the reddest rows. Ask the following in order.
Start with fit, because if your product wouldn't honestly suit what the prompt asks for, the absence isn't a content failure. Then check business relevance (does the question touch an important use case, buying criterion, or comparison), persistence (does the pattern repeat across comparable observations and more than one non-overlapping period), and evidence quality (have you read the answers and checked the sources and collection status). Last, ask whether the team can name a specific fix that could help a group of related prompts.
Funnel stage shouldn't decide priority on its own, since an implementation question central to adoption can matter more than a generic "best tools" prompt. A simple set of tiers keeps the backlog readable.
| Tier | Rule | Action |
|---|---|---|
| Fix first | Strong fit, repeated loss, verified answer and source evidence, and a specific feasible fix | Assign an owner and write the brief |
| Investigate | Important prompt, but thin data, mixed results, or an unclear cause | Collect or read more before producing anything |
| Monitor | Low relevance, weak fit, an isolated loss, or no defensible fix | Keep it visible, but out of the writing queue |
| Escalate outside content | A product-fit issue, or wrong third-party information that an article can't solve | Route to the product, technical, or communications owner |
This is a qualitative rubric I'm proposing, not a validated model or a DeepSmith priority score.
Review a rolling 30-day view and compare recent weekly slices that don't overlap, which helps separate a one-off from a pattern. It also helps to know which AI visibility metrics and KPIs you're actually using to decide.
You're done when a stakeholder can see why item one outranks item two without a composite score nobody can explain. A 55% loss rate can be an immediate candidate if the requirement matters, your product fits, and a cited source exposes a fixable gap, while a 90% loss rate on a use case you don't support can stay low priority.
Pro tip: Group related losses into one content problem when they need the same substantive answer. Avoid commissioning a thin page for every prompt variation.
Turn the diagnosis into an evidence-backed brief
A useful brief names the missing answer or evidence, and not just a topic to write about. At a minimum it should include these items.
- The prompt or cluster, the target engines, and why the requirement matters to the business
- Baseline counts, rates, and the reporting window
- Answer excerpts and cited-page records, plus your diagnosis, labeled a hypothesis where it is one
- The existing page to update, or the reason a new page is justified
- The exact factual question the content must answer, and the evidence your team can back up
- Claims the brand can't make, the owner, the change date, and the retest condition
Match the fix to the diagnosis. A coverage gap calls for a substantive answer, ideally on an existing page that's the right home. An evidence gap calls for original, checkable details that generic summaries don't have. A description error means correcting your authoritative page and finding the outdated sources. A technical gap means checking crawling, indexing, and snippet eligibility before you produce anything. A third-party source gap means correcting inaccurate listings or earning independent coverage, since a brand-owned article can't stand in for it.
Google's current guidance favors unique, expert-led, helpful content and warns against creating separate pages for every query variation. Its technical checks include crawl access, internal links, important information in text, and structured data that matches the visible page, and meeting them doesn't guarantee inclusion. For ChatGPT Search, OpenAI says access by OAI-Searchbot matters for inclusion, which is a check for one engine and not a rule for every platform. The Google guidance on AI features is worth reading directly.
This is the third place DeepSmith fits. Opportunity Agents work from tracked-prompt gaps, including getting cited for a tracked prompt, turning mentions into citations, taking a competitor's citations, and fixing how AI describes you. Every idea carries the data point that justifies it, and relevant ideas can move into New Ideas and then Planned Content in Content Studio. That helps you turn evidence into a backlog item you can defend, though it doesn't prove the cause of a loss or guarantee the finished article will win citations. The process is covered in turning AI visibility gaps into a content production backlog.
You're done when a writer or technical owner can start without inventing the missing claim or guessing which page to change. The usual errors are copying a competitor's wording and promising that schema or a particular paragraph length will force a citation, since Google says no special schema or AI text file is required.
Retest the same cohort and report the change honestly
Measure whether the pattern improved under comparable conditions, and keep that separate from claims about cause or revenue. Before you change anything, save the baseline and note the date of the intervention. After the change goes out, check these things.
Confirm the change is live and reachable, and check the relevant technical status instead of assuming publishing equals discovery. Reuse the same prompt text, engine, language, and setup where you can, compare periods of equal length, and show the valid answer counts. Keep newly added prompts apart from the fixed set, report mention, citation, competitor-only loss, and recommendation changes one at a time, and note any other material change such as a product launch or an engine change. Where it helps, watch similar prompts you didn't touch as context, without calling them a randomized control group.
An early signal is the competitor moving from the only brand named to appearing alongside yours, which isn't a sale. Don't promise a number of days until a citation change should show up, because discovery, indexing, and engine behavior all vary.
You're done when your report says what changed, on which prompts and engines, over which periods, with what counts, and what is still unknown. The usual errors are treating one good answer as proof and crediting every gain after publishing to the new content.
What to do next
Pick the single highest-priority loss that has evidence behind it, and let your prompt analytics and prompt performance tracking feed one concrete fix, not a pile of new articles. Save the baseline, and put the next comparable review on the calendar.
You can start a 7-day DeepSmith free trial to look at the answers and cited pages behind your own prompt-level gaps.
Related guides
- How to Organize Tracked Prompts by Topic, Funnel Stage and Persona: Structure a prompt list into groups by topic, funnel stage and persona so reports answer real questions.
- Why Prompt Tracking Data Goes Unused and How to Fix It: Find out why prompt data sits in dashboards and fix it with owners, a review cadence and a link to content tasks.
- How to Check Whether Your Articles Are Ready for AEO: Run a short, repeatable check on one article covering answer-first structure, clear headings, entity and schema markup, and source citations, then fix each gap you find before republishing.
- Do Traditional Content Optimizers Help With AEO?: Judge whether traditional content optimizers help AEO and where they fall short.
- How Google AI Overviews Tracking Works and Why Position Is Hard to Pin Down: Understand how AI Overviews are tracked and why position is hard to pin down.
