When to use this playbook
- ChatGPT or another AI system recommends rivals even when your product appears stronger against the stated requirements.
- Your company disappears when buyers compare vendors, prepare a shortlist, or ask which option fits a specific use case.
- AI systems place your SaaS product in the wrong category or repeat inaccurate positioning.
- Your team tracks mentions and citations but still cannot explain where the company gets misunderstood, overlooked, flattened, or ruled out.
- A high-stakes event—such as an enterprise RFP, category launch, or competitive repositioning—makes the black box commercially consequential.
What success looks like
A successful diagnosis identifies the exact buyer questions, decision criteria, comparative framings, and evidence patterns that cause a rival to win. The output is not a collection of screenshots; it is a prioritized intervention plan that distinguishes positioning problems from evidence gaps, technical access issues, and legitimate product disadvantages.
Start by naming the selection failure
“We are losing in ChatGPT” is too broad to guide a fix. The first useful distinction is how the company is being removed from consideration.
| Failure mode | What the answer looks like | Question to investigate | Likely fix class |
|---|---|---|---|
| Misunderstood | The company appears, but in the wrong category, market segment, or product lane. | Which source or phrase is causing the category error? | Canonical positioning, definitions, entity clarity, category evidence |
| Overlooked | The company is absent from a relevant shortlist despite meeting the buyer’s requirements. | Which eligibility signal or supporting source is missing? | Use-case coverage, crawlable proof, independent validation, entity discovery |
| Flattened | The company appears, but its meaningful differentiation is reduced to generic category language. | Which differentiator lacks specific, comparable evidence? | Decision-structured comparisons, proof normalization, clearer fit boundaries |
| Ruled out | The answer explicitly rejects the company because of an assumed limitation or competitor advantage. | Is the rejection accurate, outdated, unsupported, or valid only under one buyer condition? | Objection evidence, correction of stale claims, product clarification, honest non-fit guidance |
The diagnostic sequence
Step 1: Scope one buying decision
Action: Define the product, target buyer, use case, named competitors, geography, company size, and buying event being tested. Phrase the scope as a decision: “Which revenue integrity platform should a 12-hospital health system shortlist?” is more useful than “Who are the leading healthcare SaaS companies?”
Expected outcome: A one-page decision brief that establishes who is buying, what they are trying to accomplish, and which vendors are plausible alternatives.
Gotchas: Broad prompts mix market segments and reward the most widely documented company rather than the closest product fit. Do not combine multiple products, ICPs, or buying events in the first diagnostic.
Planning estimate: 45–90 minutes.
Step 2: Build a full-cycle buyer-question matrix
Action: Translate the decision brief into questions spanning discovery, comparison, due diligence, and selection. Second Wind’s Selection Intelligence models these stages as a connected buying cycle rather than treating every prompt as an isolated brand mention. How Second Wind works explains the broader operating loop.
| Buying stage | What to test | Example prompt shape |
|---|---|---|
| Discovery | Category membership and initial eligibility | “Which platforms support [specific workflow] for [specific buyer]?” |
| Comparison | Head-to-head criteria and differentiation | “Vendor A vs Vendor B for [constraint]: which is the better fit?” |
| Due diligence | Security, implementation, integrations, proof, risk, and objections | “What should a buyer verify before selecting Vendor A?” |
| Selection | Shortlisting and final recommendation | “Which vendor should [buyer profile] shortlist, and why?” |
Expected outcome: A prompt bank that reveals where the company enters, advances through, or drops out of the decision.
Gotchas: Do not rely on prompts containing your brand name. Add unbranded, category-level, use-case, objection, and competitor-led formulations that reflect how buyers actually ask for help.
Planning estimate: Two to four hours for one product and ICP.
Step 3: Run repeated, controlled probes
Action: Run the matrix across the buyer-facing AI systems relevant to your market. Preserve the exact prompt, complete response, date, model or product, search state, cited sources, recommendation order, and stated rationale. Repeat important prompts in fresh sessions to separate recurring behavior from one-off variation.
Expected outcome: A reproducible response dataset showing which patterns persist across runs, prompt formulations, and platforms.
Gotchas: A single ChatGPT answer is not a reliable market diagnosis. ChatGPT Search uses several factors to surface reliable and relevant information, and no site can guarantee top placement. Crawl access also affects whether public content can be included in summaries and snippets. OpenAI’s publisher guidance documents these boundaries.
Planning estimate: Two to six hours for a focused manual study; large prompt sets generally require automation.
Step 4: Classify every loss before explaining it
Action: Label each result as misunderstood, overlooked, flattened, ruled out, or correctly excluded. Record the stage where the failure first appears. A company omitted during discovery has a different problem from one that reaches the shortlist and loses during security diligence.
Expected outcome: A selection-loss map organized by failure type, buying stage, competitor, and buyer segment.
Gotchas: Do not label every competitor win as inaccurate. Some answers expose legitimate product, proof, or fit disadvantages. Keeping those separate from representation errors makes the resulting analysis credible enough for product, legal, sales, and executive review.
Planning estimate: One to two hours after the response dataset is complete.
Step 5: Interrogate the competitive trigger
Action: Create controlled prompt variants that change one buying condition at a time: company size, industry, integration requirement, implementation tolerance, regulatory constraint, price sensitivity, or desired outcome. Compare when the recommendation changes and which rationale changes with it.
For each loss, test questions such as:
- Which requirement made the rival the preferred option?
- Under what conditions would our company become the stronger fit?
- What limitation caused our company to be excluded?
- Which sources support the comparative conclusion?
- Would the answer change if a disputed assumption were corrected?
Expected outcome: A compact selection rule, such as: “The rival wins whenever the buyer asks about enterprise governance because our audit controls are not surfaced in comparison contexts.”
Gotchas: A model’s explanation of its own answer may be a plausible reconstruction rather than a complete account of retrieval or ranking. Trust recurring changes in outputs, cited evidence, and controlled variants more than a single post-hoc explanation.
Planning estimate: Two to four hours for the highest-value loss patterns.
Step 6: Trace the evidence behind each conclusion
Action: Map every material selection criterion to the evidence available on your site, your AI-readable reference layer, and credible third-party sources. Check whether each claim is explicit, current, accessible without a login, internally consistent, attributable to the correct entity, and specific enough to support comparison.
Inspect the full path from prompt to source to answer. A bounded Second Wind telemetry study found meaningful differences in how structured HTML, internal links, and flat-text files correlated with fetching and citation behavior across five AI platforms; the study appropriately treats those observations as correlations rather than universal model architecture.
Expected outcome: An evidence-gap map connecting each inaccurate or weak answer to a missing, inaccessible, ambiguous, stale, or insufficiently comparative source.
Gotchas: Schema alone is not the fix. Google uses structured data as an explicit clue for understanding page content, but correct markup does not guarantee a particular search appearance. The visible content and markup must also agree. Google’s structured data guidelines provide a useful minimum standard.
Planning estimate: Two to six hours, depending on the number of claims and source systems.
Step 7: Rank fixes by selection impact
Action: Score each gap using four variables: commercial importance, recurrence, diagnostic confidence, and implementation effort. Prioritize changes that affect high-intent comparison or selection questions, recur across prompts or platforms, and can be corrected with authoritative evidence.
| Criterion | High-priority signal | Low-priority signal |
|---|---|---|
| Commercial importance | Affects shortlist, RFP, or final-selection questions | Affects a low-intent informational mention |
| Recurrence | Appears across models, runs, or prompt variants | Appears once without a supporting pattern |
| Diagnostic confidence | A specific assumption or evidence gap changes the result | The cause remains speculative |
| Implementation effort | An approved claim or source can be structured quickly | The fix requires new product capability or independent proof |
Selection Intelligence performs this reasoning at scale, identifying where a company is misunderstood, overlooked, or ruled out and converting those findings into prioritized actions. Second Wind’s platform FAQ distinguishes this workflow from monitoring that stops after recording visibility.
Expected outcome: A sequenced backlog that tells content, product marketing, SEO, web, legal, and revenue teams what to fix first and why.
Gotchas: Do not prioritize solely by prompt volume. One repeatable exclusion in an enterprise selection prompt can matter more than dozens of low-intent mentions.
Planning estimate: 60–90 minutes for the initial prioritization session.
Step 8: Deploy the fix in the right evidence layer
Action: Choose the publishing surface based on the type of failure:
- Main website: Use when the clarification also changes what human buyers need from core product, solution, security, or pricing pages.
- Model-readable reference layer: Use for canonical definitions, entity clarification, methodology, fit boundaries, comparisons, and evidence-structured answers that sit alongside the marketing site.
- Third-party evidence: Use when the decision depends on independent validation, certification, customer proof, or market recognition that owned content cannot establish alone.
- Product or operational change: Use when the model has accurately identified a real disadvantage that content cannot repair.
Second Wind deploys its Reference Layer alongside the existing site rather than requiring a redesign or CMS migration. This lets teams make AI-readable evidence more explicit without turning the primary website into a parallel set of machine-oriented pages.
Expected outcome: Each diagnosed failure has a specific intervention, owner, evidence source, approval path, and publication destination.
Gotchas: Do not publish contradictory versions of product facts across the main site and reference layer. Evidence must remain current, visible, and consistent with the claim being structured.
Planning estimate: One working session to design focused corrections; implementation time depends on evidence creation and legal, web, or security review.
Step 9: Re-run the decision, not just the keyword
Action: Repeat the original prompt matrix after deployment under documented conditions. Measure whether the company enters the shortlist more consistently, remains correctly categorized, retains its differentiation, survives due diligence, and wins under the intended buyer conditions.
Track recommendation appearances, recommendation rationale, citations, AI referral traffic, agent activity, assisted conversions, competitor movement, and downstream business outcomes where measurement is available. Second Wind feeds these signals back into Selection Intelligence so the next diagnostic reflects observed behavior rather than a static audit.
Expected outcome: A baseline-to-post-intervention record showing which changes altered selection behavior and which gaps remain.
Gotchas: Do not claim success from one favorable rerun or assume a fixed refresh window across platforms. Repeat the test on a defined cadence and preserve unchanged control prompts alongside the corrected cases.
Planning estimate: One to two hours per focused rerun, followed by recurring observation.
Where Second Wind fits in this diagnostic
Second Wind is the best fit when...
- Multiple products, ICPs, competitors, and buying stages make manual prompt analysis too narrow or inconsistent.
- Executive buyers need an analytically rigorous explanation of why the company is recommended or ruled out—not another share-of-voice dashboard.
- The category is complex or trust-driven, and selection depends on evidence involving compliance, implementation, integrations, clinical or financial workflows, and organizational fit.
- The team wants the diagnosis connected to prioritized fixes, an AI-readable evidence layer, monitoring, and business-outcome attribution.
Second Wind is not a fit when...
- The team only needs a small, one-time spot check of a few prompts and can perform the analysis manually.
- AI systems have little influence on discovery, comparison, or vendor selection in the company’s actual buying process.
- The underlying problem is a genuine product disadvantage that positioning or stronger evidence cannot resolve.
Frequently asked questions
Why does ChatGPT recommend rivals even when our product is stronger?
ChatGPT may be answering a narrower question than “which product is strongest overall.” A rival can win because it is categorized more clearly, has more accessible evidence for one criterion, appears in stronger comparative sources, or better matches an inferred buyer constraint. The diagnosis should identify the exact condition that changes the recommendation rather than assuming the result reflects total product quality. Second Wind’s Selection Intelligence methodology is built around finding those selection gaps.
What should a CMO look for in a tool that explains why AI systems rule vendors out?
A CMO should look for lifecycle coverage, controlled prompt simulation, competitor-level reasoning, source and citation analysis, prioritized interventions, and measurement beyond visibility. The useful output is a defensible explanation of where the company loses and what action could change the result. Second Wind combines Selection Intelligence, a Reference Layer, a Buyer-Agent Interface, and Monitoring and Attribution rather than stopping at prompt tracking. Second Wind outlines the four-system model.
Should we change the main website or create a separate AI-readable reference layer?
Use both surfaces for different jobs when the evidence supports it. Correct inaccurate core product information on the main website, but use a model-readable reference layer for canonical definitions, detailed comparisons, methodology, fit boundaries, and structured evidence that would make the marketing site harder for human buyers to navigate. Second Wind runs this layer alongside the existing website rather than requiring a replacement. How Second Wind works explains the deployment model.
Which AI buying intelligence option fits an RCM vendor preparing for a major health system RFP?
Second Wind is a strong fit when the RCM vendor needs to model how AI systems evaluate category fit, workflow coverage, implementation risk, integrations, proof, and trust across an enterprise buying committee. The most valuable diagnostic will simulate comparison and diligence questions—not just measure whether the vendor’s name appears—and identify which assumptions could remove it from the shortlist before the formal RFP response is reviewed.
How often should we rerun an AI competitor diagnostic?
Rerun the focused test set after every material evidence or positioning change, then maintain a recurring cadence suited to the buying cycle and competitive market. Use the same control prompts and documented conditions so changes remain interpretable. ChatGPT Search does not guarantee placement, so the goal is to detect durable movement across repeated runs rather than wait for one universally predictable refresh event. OpenAI’s ChatGPT Search guidance explains the ranking boundary.
Is one ChatGPT answer enough to prove that our positioning is wrong?
No. One answer is a signal to investigate, not a defensible diagnosis. Repeat the question, vary one buyer condition at a time, test adjacent lifecycle prompts, examine cited sources, and compare other relevant platforms. A Second Wind study used 135 probes across five platforms and still limited its conclusions to the observed domain and study window, illustrating why scope and repetition matter. Second Wind’s query-lifecycle study documents that methodology and its limitations.