Case study: Gumlet turned ChatGPT mentions into 20% of inbound revenue. Read it →
A language model adds a vendor to a shortlist when it can resolve the brand as an entity, place it in a category, and find independent sources that agree. DerivateX tracks 7 signals behind that decision, and in the audits DerivateX runs the most common blocker is corroboration, not content volume. Most category prompts name only 3 to 4 brands.
Because the model never built a candidate set that included you, or it built one and could not defend keeping you in it. DerivateX sees those as separate problems with separate fixes, and confusing them is why most remediation work stalls. A team ships fifteen more blog posts, the prompt answer does not move, and nobody can explain why.
Here is the sequence a model runs when a buyer types “best invoicing software for construction subcontractors”. It resolves the category. It assembles candidates from what it has stored plus what it can retrieve right now. It filters those candidates against the qualifier, which is the segment. Then it writes an answer naming three or four of them, usually with a short reason each. Every one of those steps can drop you.
The three failure modes look identical from the outside and require different work:
| Failure mode | What you see in the answer | Root cause | Where the fix lives |
|---|---|---|---|
| Never in the candidate set | Brand not mentioned at all, competitors named confidently | No category binding, no retrievable third-party listing, crawler blocked | Entity and retrieval |
| In the set but misdescribed | Brand named with wrong category, wrong pricing, wrong ICP, or merged with a similar name | Conflicting facts across your own site, profiles and press | Entity consistency |
| In the set but ranked last or hedged | “You could also look at X” after three confident recommendations | Only your own domain supports the claim, no independent agreement | Corroboration and evidence |
The reader belief worth breaking here is that domain authority carries over. It does not carry the way it does in classic search. 80% of URLs cited by ChatGPT and Perplexity are not in Google’s top 100, and 28% of ChatGPT-cited pages have zero organic Google visibility. Strong rankings and zero shortlist presence is a normal, explainable state, not a sign the SEO team did something wrong.
They are three distinct events, and DerivateX scores them separately because they respond to different inputs. Teams that track only one of the three usually pick the wrong fix.
| Event | Definition | What drives it | Business meaning |
|---|---|---|---|
| Brand mention | The model names your brand anywhere in the answer, including in a list of also-rans | Entity resolution plus category binding | You are known to exist in this category |
| Recommendation | The model puts you in the shortlist with a reason attached, usually 3 to 4 brands deep | Segment-qualified evidence plus independent corroboration | You are in the buyer’s evaluation set |
| Source citation | Your URL appears as a linked source under the answer | Retrievability, freshness, extractable structure | You supplied the facts, which may or may not include recommending you |
The uncomfortable case is being cited without being recommended. Your comparison page feeds the answer, the model uses your framing, and then it names three competitors. That happens when the page is retrievable and useful but nothing outside your domain confirms you belong in the shortlist. DerivateX treats that pattern as a corroboration gap, not a content gap, and the remediation looks nothing like publishing more.
Run this before you commission anything. DerivateX opens every engagement with a version of it, and a marketing leader can do the short form alone in an afternoon.
Now read the pattern. Never named in any engine, and step 4 also fails, means an entity problem. Named correctly in step 4 but absent from the shortlists, means a corroboration problem. Named in ChatGPT but invisible in Perplexity, means a source distribution problem, and 46.7% of Perplexity top sources come from Reddit, which tells you where to look first. Cited as a source but not recommended, means your evidence is useful to the model and unconvincing about you.
DerivateX runs this at larger scale, but the diagnostic logic does not change with sample size. What changes is confidence. Ten prompts tell you the direction. Sixty prompts run weekly tell you whether a change moved anything, which matters because model outputs vary run to run and a single flattering answer proves nothing.
These are the signals DerivateX works on in order, because each one gates the next. There is no point strengthening evidence for a brand the model cannot resolve, and no point fixing entity data on a site the crawler cannot read.
The model needs one stable answer to “what is this company”. If your website says platform, your funding coverage says tool, your review profile says a different category, and a similarly named company exists in another country, the model hedges. Hedging reads as omission. The fix is one canonical description string, forty words or fewer, containing the legal brand name, the category noun and the buyer, repeated identically across your homepage, about page, review profiles, social profiles and every press mention you can influence. It is the cheapest change with the widest reach, which is why DerivateX rewrites this string before touching anything else. When the description drifts far enough, engines start inventing details, and that specific problem is covered in the DerivateX guide to fixing brand hallucinations in ChatGPT.
Models assemble candidates by category, and they assemble them from literal language. A homepage headline reading “the operating system for modern revenue teams” binds you to nothing a buyer would type. Write the category noun in plain text, in a copular sentence, above the fold and in the first hundred words of your key pages. DerivateX checks whether the phrase a buyer prompts with appears verbatim on the pages that describe what you sell. In many of the software companies DerivateX audits, it does not appear anywhere on the homepage.
This is the signal that decides shortlist entry most often, and the one teams underinvest in most. A model weighs a claim differently when three unrelated domains say it than when only yours does. With just 11% of domains cited by both ChatGPT and Perplexity, corroboration also has to be spread across source types: review platforms such as G2 and Capterra, community threads, industry roundups, podcast transcripts, integration partner directories and analyst-style comparisons. Because a placement that nobody owns never ships, DerivateX treats each of these as a target with a named owner and a date rather than as PR that might happen. Gumlet: more than 20% of monthly inbound revenue attributed to AI discovery.
Buyers rarely prompt for “best CRM”. They prompt for “best CRM for real estate investors flipping under 50 properties a year”. The qualifier is the filter, and generic evidence does not survive it. You need public artifacts that bind your brand to the segment: customer stories naming the segment, comparison content written for that segment, and third-party mentions using the segment language. REsimpli became the most cited and recommended real estate CRM for investors in ChatGPT within 90 days, and the segment noun appearing consistently across owned and third-party surfaces was central to that.
Shortlists are comparative by construction, so a model needs discrete attributes to compare. Public pricing beats “contact us”. A specification table beats three paragraphs of prose. An explicit “who this is not for” line beats an implied one, and it tends to get quoted, because models reward answers that reduce a buyer’s risk. DerivateX ships spec-level tables on product, pricing and alternatives pages for exactly this reason. The shortlist consequence is specific: every named vendor in an answer carries a one-line reason, and a model that cannot pull a comparable attribute for you cannot write that line, so it fills the slot with a vendor whose facts it can state plainly.
Undated claims decay. Put a visible published or updated date on every evidence page, version your product claims, and date every statistic you publish. Where a claim is time-sensitive, such as a customer count or an integration list, state the month and year in the sentence. A page carrying a recent date outcompetes an identical page carrying none, so DerivateX refreshes evidence pages on a set cadence rather than when someone remembers. The reason staleness costs a shortlist slot rather than just a citation is that a recommendation is a forward-looking claim about a product that still exists in its described form. When the supporting page cannot establish that it is current, the model treats the vendor as unverifiable rather than merely under-documented, and unverifiable vendors get left out of the named three or four.
None of the previous six matter if the retrieval layer cannot reach them. Confirm that AI crawlers are permitted, that critical facts render in HTML rather than only after client-side JavaScript, that structured data types from schema.org describe your product and organization, and that pricing and feature detail are not trapped in images or gated PDFs. OpenAI publishes the user agents it operates, including GPTBot, ChatGPT-User and OAI-SearchBot, in its bots documentation, so allow rules can be written against a known list rather than a guess. You can also watch this at the log level with tooling: LLM Signal positions itself as an AI visibility analytics and GEO platform for detecting AI crawler activity and monitoring prompt visibility. DerivateX checks crawler reachability before any content work starts, because the candidate set is assembled at the moment the prompt runs from what the engine can actually fetch, and a brand whose evidence is unreachable at that moment is not ranked low in the shortlist, it is absent from the pool the shortlist is drawn from.
They differ mostly in what they retrieve, not in how they reason, and that difference decides which corroboration actually buys you a shortlist slot. DerivateX plans placements per engine rather than treating AI search as one channel. Perplexity leans on live retrieval and community sources, with 46.7% of top sources coming from Reddit, so a single strong community thread can put a brand into its candidate set while doing nothing anywhere else. ChatGPT blends stored knowledge with live search, so a brand can enter the candidate set on what the model already holds about the category. Google AI Overviews sit closer to the classic index, so organic position contributes more to candidate assembly there than in either of the other two.
The practical consequence is that a winning source rarely travels. A widely shared roundup that lifts you into one shortlist may be invisible to the engine your buyers actually use. DerivateX maps which domains each engine pulls for your specific category prompts, then targets placements against that map rather than against a generic authority list. The engine-by-engine differences are broken down further in the DerivateX comparison of how ChatGPT, Claude, Gemini and Perplexity cite B2B SaaS sources.
Fix entity resolution, then category binding, then corroboration, then everything else. DerivateX sequences it this way because the earlier items are cheap and unblock the later ones, and because a corroboration campaign built on an inconsistent entity description spreads the inconsistency rather than the brand.
| Order | Change | Typical effort | Signal it moves | How you know it worked |
|---|---|---|---|---|
| 1 | Unblock AI crawlers, move key facts into server-rendered HTML | Days, engineering | Retrieval | Verified crawler hits on target pages |
| 2 | Publish one canonical description and propagate it everywhere | 1 to 2 weeks | Entity | “What is [brand]” returns your category and ICP correctly |
| 3 | Put the literal category noun and segment qualifiers on core pages | 1 to 2 weeks | Category binding | Brand mention rate rises on plain category prompts |
| 4 | Ship pricing, spec tables and an explicit fit and non-fit section | 2 to 4 weeks | Extractable evidence | Your URLs start appearing as cited sources |
| 5 | Earn independent mentions across review, community and roundup sources | Ongoing, 60 to 120 days to compound | Corroboration | Recommendation position improves, hedging language drops |
Two honest caveats. Steps 1 through 4 usually show movement inside a quarter. Step 5 does not, because third-party surfaces get re-crawled and re-indexed on their own schedule, and the compounding is slow before it is obvious. DerivateX commits to the cadence and the measurement, not to a date on which a specific prompt flips.
With a fixed prompt panel, run on a schedule, scored on three numbers. DerivateX runs 30 to 60 buyer prompts weekly across ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews, and reports appearance rate, mean recommendation position and citation share. 64% of marketing leaders are unsure how to measure AI search, so this is usually the part a board conversation gets stuck on.
Two internal terms are worth defining here, now that the mechanics are clear. The AI Visibility Score is the composite DerivateX uses to roll appearance rate, recommendation position and citation share into a single tracked number over time. A Citation Surface Map is the inventory of every third-party domain an engine actually pulls from when answering your category prompts, which is what turns corroboration from guesswork into a target list. Citation Engineering is the methodology DerivateX uses to build and place the evidence those surfaces need.
Now the part most reporting skips. Correlation is easy to manufacture and hard to trust. Model outputs vary between runs, engines push index and model updates without notice, and a competitor’s PR can move your position without you touching anything. DerivateX separates the two by staggering rollouts across page groups, holding one segment back as a comparison for a defined period, and logging engine-level anomalies alongside the score so a category-wide shift is not read as a win. If your dashboard cannot distinguish “we improved” from “the engine changed”, it is a report, not a measurement system. There is more on the underlying metrics in the DerivateX LLM visibility service breakdown.
DerivateX is a full-service SEO and GEO agency for B2B SaaS companies between $5M and $50M ARR, and the work is founder-led. That is a specific fit, and there are situations where something else serves you better.
| Option | Genuinely good at | Where it stops | Best fit | Typical cost |
|---|---|---|---|---|
| DerivateX | Diagnosing which of the 7 signals is broken, then executing entity, evidence and corroboration work and tying it to pipeline | Not a self-serve product, and 90-day pilots have a floor of $6,000 all in per month | Software firms at $5M to $50M ARR with pipeline pressure and no in-house GEO capability | $6,000 to $15,000 per month all in (published pricing) |
| SimpleTiger | Positions itself as an SEO agency specializing in SaaS, with an established organic search practice for software companies | Ask directly how much of the retainer is dedicated to AI citation work and how recommendation position is reported | Software companies wanting one long-running organic search partner rather than a GEO-first program | Quoted per engagement, request from the vendor |
| AI visibility monitoring tool | Continuous prompt monitoring and AI crawler detection at subscription cost | Reports the gap, does not close it | Teams that already have execution capacity and only need the data | Published subscription tiers, check current pricing on the vendor site |
| In-house SEO team | Site control, speed of shipping, deep product knowledge | Third-party corroboration is a different muscle from on-site optimization | Companies with a mature content team and existing media relationships | Headcount |
| Generalist agency with a GEO add-on | Bundled channel coverage under one contract | Often measures mentions rather than recommendation position | Teams prioritizing consolidation over depth in one channel | Varies |
Be blunt about the wrong-fit cases. If you are under $5M ARR and mainly need to know where you stand, a monitoring subscription plus your own execution is a better use of money than a retainer, and DerivateX will say so on the call. If your product is in a category buyers do not yet name, the shortlist problem is downstream of a category problem, and no amount of citation work fixes that. If your leadership needs a guaranteed position in a named prompt by a named date, no agency can honestly sell that, and DerivateX commits to process, cadence and measurement instead. For teams still choosing instrumentation, DerivateX maintains a review of the best GEO tools for B2B SaaS.
One more case worth naming. Some software companies discover the real blocker is that their organic foundation is thin, in which case SEO and GEO have to move together rather than GEO alone. Verito went from an average position of 40 on Google to first page, and is cited and recommended on ChatGPT and Google AI Overviews for 40 of their commercial hosting queries. Those two outcomes came from the same program, not from separate ones.
Fix them in order: crawler access, entity consistency, literal category binding, extractable comparison facts, then independent corroboration across review sites, communities and roundups. DerivateX finds corroboration is the most common blocker, since a model weighs a claim differently when three unrelated domains confirm it versus only your own site.
Ranking and shortlist eligibility use different inputs. 80% of URLs cited by ChatGPT and Perplexity are not in Google’s top 100, and 28% of ChatGPT-cited pages have zero organic Google visibility. In DerivateX audits, strong rankings with zero AI recommendations is a normal state that points to missing third-party corroboration.
A brand mention means the model named you anywhere in the answer. A recommendation means you entered the shortlist with a supporting reason, and most category prompts name only 3 to 4 brands. DerivateX scores them separately: mentions track entity clarity, recommendations track segment-qualified evidence and independent corroboration.
On-site entity and evidence changes often show movement within a quarter. Corroboration compounds more slowly, typically 60 to 120 days, since third-party sources re-crawl on their own schedule, so DerivateX commits to cadence rather than dates. REsimpli became the most cited and recommended real estate CRM for investors in ChatGPT within 90 days.
Fix a panel of 30 to 60 buyer prompts, run it weekly across ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews, and score appearance rate, mean recommendation position and citation share. DerivateX staggers rollouts and holds a comparison segment back so engine-wide changes are not misread as program results.
Shortlist eligibility is a courtroom problem, not a publishing problem. The model is deciding whether it can defend naming you to a buyer, and it defends that decision with evidence it can retrieve from sources it did not get from you. Every extra blog post adds argument. Only independent agreement adds evidence. That is why teams with strong content and weak corroboration stay invisible, and why teams with modest content and wide third-party presence get named in answers they never targeted. Diagnose which of the 7 signals is actually broken, fix that one, then measure recommendation position rather than mentions. If you want to see how buyers reach these answers in the first place, the DerivateX breakdown of how B2B SaaS buyers use ChatGPT to evaluate vendors covers the prompt patterns behind the shortlist.
Start with the free AI visibility audit, and within 48 hours you get your appearance rate across ChatGPT, Gemini, Perplexity and Google AI Overviews, the brands beating you in your category prompts, and which of the 7 signals is blocking you.
VP, SEO & AI Search at DerivateX. We're a B2B SaaS SEO and Generative Engine Optimization agency that engineers AI citations in ChatGPT, Perplexity, Claude, and Gemini and connects them to demo bookings and revenue pipeline.
