Case study: Gumlet turned ChatGPT mentions into 20% of inbound revenue. Read it →
AI search vs Google rankings is not one contest: ranking first on Google and being recommended by ChatGPT are decided by different systems. DerivateX treats the gap as three separate failures, no mention, mentioned but ranked last, or a competitor’s page cited. With only 3 to 4 brands named per AI category query, position one buys nothing.
Because ranking and recommendation are two separate selection events. Google scores documents against a query and returns a list. A language model does something else: it retrieves a set of passages, builds a candidate set of brands from what it retrieved plus what it already holds from training, then writes an ordered recommendation. Your page can win step one and never enter step two.
The overlap between those two sets is thinner than most SEO teams expect. 80% of URLs cited by ChatGPT and Perplexity are not in Google’s top 100. 28% of ChatGPT-cited pages have zero organic Google visibility at all. Only 11% of domains are cited by both ChatGPT and Perplexity, which tells you the engines do not even agree with each other, let alone with Google. DerivateX runs this comparison across engines for every client at kickoff, and the pattern holds in almost every B2B SaaS category we have tested.
This matters commercially because the front door moved. 40% of Google queries now show Google AI Overviews, and click-through rate drops 61% when they appear. 73% of B2B sites lost significant traffic between 2024 and 2025. If you run SEO for a software company, none of that was caused by your work, and better rankings would not have prevented it. The traffic did not go to a competitor who out-ranked you. It went into an answer that never needed a click.
There is a payoff on the other side. Visitors who arrive from AI-sourced discovery convert at 4.4x the rate of other channels, because the assistant already did the qualifying, which means fewer sessions but better ones. Our benchmark of ChatGPT against Google AI Overviews breaks down where the two surfaces diverge on source selection.
Run the same buyer prompt across four engines and record three separate things, not one. DerivateX treats these as three independent columns because they fail for different reasons and respond to different work. Most teams collapse them into a single yes-or-no and then fix the wrong thing for six months.
Brand mention is whether your name appears anywhere in the answer. Recommendation position is where you sit in the ordered list and what qualifier is attached, because “also worth considering if budget is tight” is not the same outcome as “best for teams under 200 seats”. Source citation is whose URL is footnoted underneath. You can be recommended without being cited, and cited without being recommended.
| What you see in the answer | What it actually means | Root cause | First fix |
|---|---|---|---|
| Not mentioned at all, competitors named | You are not in the candidate set | Entity resolution failure, or no corroboration outside your own domain | Fix entity consistency, then build third-party evidence |
| Mentioned last, hedged qualifier | You are in the set but rank low on evidence weight | Thin or unverifiable proof on the specific attribute the prompt asked about | Publish machine-readable proof for that one attribute |
| Your page is cited, a competitor is recommended | Your content is background material, not vendor evidence | Educational content with no product claim inside the retrievable passage | Add a self-contained product claim to the chunk that gets pulled |
| Mentioned but described wrongly | Stale or conflicting facts in the model’s grounding | Outdated pricing, old positioning, renamed product, conflicting directory entries | Correct the source records, not just your homepage |
| Mentioned in ChatGPT, absent in Perplexity | Source-mix mismatch, not a brand problem | Perplexity weights community sources heavily, and 46.7% of its top sources come from Reddit | Work the community surface, not more owned pages |
Give this thirty minutes. Pick 25 to 40 prompts a real buyer would type, run them through ChatGPT, Gemini, Perplexity and Claude, and fill the grid. You will usually find that one row dominates. That row is your actual bottleneck, and the other six reasons below are noise for you right now. If you want a sense of how buyers phrase those prompts in the first place, our write-up on how B2B SaaS buyers use ChatGPT to evaluate vendors has the patterns we see most.
These are the failure modes DerivateX sees most often in $5M to $50M ARR B2B SaaS accounts that already rank well. They are ordered by how often they turn out to be the primary cause, not by how hard they are to fix.
Entity resolution comes before ranking of any kind. If a model cannot confidently link your brand name to a category, a product type and a buyer, it will not risk naming you in a recommendation. Ambiguity is expensive for an assistant, so it defaults to brands it can describe in one clean sentence.
This breaks in mundane ways. Your homepage calls you a “revenue intelligence platform”, your G2 category says “sales engagement”, your Crunchbase blurb says “AI for GTM teams”, and your docs call the product something else again. Four descriptions, no agreement, no confident entity. DerivateX starts almost every engagement by forcing one canonical description across owned, earned and structured sources, which is the groundwork covered in our guide to brand grounding for AI search.
Buyers type prompts. Prompts carry constraints that keywords never did: team size, stack, budget ceiling, compliance requirement, migration source. “Best CRM” is a keyword. “Best CRM for a 40-person real estate investment firm that needs skip tracing built in” is a prompt, and it produces a completely different candidate set.
Vocabulary also shifts the result more than people expect. In DerivateX’s own prompt testing, the same entity returned a 36% mention rate for prompts phrased around “B2B”, 31% for “SaaS”, and 12% for “software companies”. The brand was welded to one phrasing and simply did not resolve for the others. If your category page only ever uses one noun for what you sell, you are invisible to a third of the phrasings your buyers use. Treat this as correlation observed across our prompt panels, not a proven causal law.
Pricing sits behind a form, the integration list renders in JavaScript, the security posture lives in a PDF behind a trust portal login, and limits and quotas are answered only by sales. Every one of those is a claim the model cannot verify, so the model recommends a vendor whose equivalent claim is sitting in plain HTML.
Compare that to software firms that publish granular detail openly. HighLevel states exact per-unit email sending rates in its public HighLevel pricing and billing guide, with no form and no sales call in the way. Salesmotion’s pricing page publishes its per-seat plan price and the included verified contact credit volume, alongside its own comparison of what enterprise intelligence platforms typically charge annually. Neither company sells into the categories DerivateX works in most often, which is exactly the point: answerability is a publishing decision, not a category advantage. A model asked “how much does X cost” can quote them and cannot quote you.
Self-description carries almost no weight in a recommendation decision. Corroboration does. When three independent sources describe your product the same way, the model treats the claim as settled. When only your website says it, the claim stays a marketing statement.
This is where community and review surfaces decide outcomes. 46.7% of Perplexity’s top sources come from Reddit, which means a category thread you have never read may be doing more to shape your shortlist position than your last twenty blog posts. Review profiles, analyst directories, comparison roundups and practitioner posts all function as corroborating records. DerivateX builds these deliberately rather than hoping for them, and the method is set out in our breakdown of third-party assets for AI search.
Retrieval happens at the passage level. A 3,000-word pillar page that builds an argument across ten paragraphs is excellent for a human reader and close to useless for an engine that needs one self-contained block containing the entity, the claim and the qualifier together.
DerivateX rewrites for chunk survival, which means each section has to answer its question inside the first two sentences and name the subject explicitly rather than relying on “it” or “we”. A paragraph that says “we support SOC 2 and HIPAA” is unattributable the moment it is lifted out of the page. A paragraph that says “[Brand] supports SOC 2 Type II and HIPAA, with the audit report available without an NDA” survives the trip into an answer intact, which is the same fact carrying completely different retrieval value.
When a buyer asks “X versus Y” or “alternatives to X”, the model pulls from pages that already do that comparison. If every “best tools for [category]” listicle in your space was written by a competitor’s content team or an affiliate site that never included you, that is the corpus the recommendation is built from. You are not losing on merit, you are absent from the ballot.
Fixing this is unglamorous, and DerivateX treats it as outreach work rather than content work: get included in the roundups that already rank and already get cited, publish your own honest comparisons that name rivals and their genuine strengths, and make sure your positioning statement appears near your competitor names in text the model can retrieve. Our analysis of the signals that drive LLM vendor shortlists goes deeper on which of these surfaces carry the most weight per unit of effort.
Training data does not refresh on your release schedule. If you repositioned eighteen months ago, raised your price, dropped a product line or renamed a module, the model may still be describing the previous version of your company with total confidence. Buyers reading that answer are comparing your 2023 product against a competitor’s 2026 one.
DerivateX handles this by correcting the underlying records rather than the homepage alone: directory entries, review profiles, documentation, press coverage, and any widely-cited third-party page carrying the stale fact. The homepage is the one source the model is least likely to need, because everyone else already summarized it.
Sequence by diagnosis, not by effort. DerivateX prioritizes in this order because the earlier fixes gate the later ones: an entity that will not resolve cannot benefit from better evidence, and better evidence cannot help if nothing outside your domain repeats it.
| Diagnosis from your grid | First action | Signal to watch | Realistic time to move |
|---|---|---|---|
| No mention in any engine | Canonical entity description across owned, structured and third-party records | Mention rate on your core 25 prompts | 4 to 8 weeks |
| Mentioned, never recommended | Publish verifiable proof for the one attribute the prompt asks about | Position within the ordered list, and the qualifier attached | 6 to 10 weeks |
| Recommended in ChatGPT only | Work community and review surfaces that Perplexity weights | Cross-engine coverage, engine by engine | 8 to 12 weeks |
| Described inaccurately | Correct third-party records carrying the stale fact | Accuracy rate of the description in answers | 6 to 12 weeks, depends on source refresh |
| Cited but not recommended | Add self-contained product claims to the passages being pulled | Ratio of citations to recommendations | 4 to 8 weeks |
Citation Engineering is the methodology DerivateX uses to make language models recommend a brand on purpose rather than by accident, and it starts with a Citation Surface Map, which is a record of every place outside your own domain where your category is currently being described to an engine. The map usually surprises people. Half the sources deciding your shortlist position are pages your marketing team has never touched and does not own.
These are targets, not guarantees. Nobody controls a model’s output, and any agency telling you otherwise is selling something that does not exist. What DerivateX commits to is the cadence, the measurement and the evidence trail behind each change.
Measure recommendation behavior on a fixed prompt panel, and hold the panel constant. DerivateX tracks a locked set of buyer prompts weekly across ChatGPT, Gemini, Perplexity and Claude, logging mention, position, qualifier and cited source separately, plus the model version at time of run. Change the prompts and you have destroyed your own baseline.
The AI Visibility Score is the composite DerivateX uses to summarize that panel into one number a marketing leader can take to a board review: share of prompts where the brand is mentioned, weighted by position and by whether a brand-controlled source was cited. It is a directional index, not a precision instrument. 64% of marketing leaders say they are unsure how to measure AI search, and the reason is usually that they are trying to read it off a traffic chart instead of off the answers themselves.
Two things will corrupt your reading if you let them. Models update, sometimes silently, and a jump in your numbers the week after a model release probably belongs to the release rather than your work. Run a control set of prompts you are deliberately not working on, and compare the movement. If the control moves as much as the treated set, you learned something about the model and nothing about your campaign. That discipline is what separates a measurement program from a dashboard, and it is the axis on which we assess tools in our comparison of LLM visibility trackers for B2B SaaS.
Tie it to pipeline at the end. Ask every inbound lead where they first heard of you, add an open-text field, and count the answers that name an assistant. With 200M+ weekly ChatGPT Search users, that field fills faster than most teams expect. Gumlet now attributes more than 20% of monthly inbound revenue to AI discovery, which is the number that ends the internal debate about whether this channel is real.
Often, and DerivateX would rather say it here than after you have signed something. If your diagnostic grid shows a single failure mode and it is a technical one on your own site, buy software and keep your money. Plenty of software companies in this bracket have an on-site structure problem and nothing else.
Rank Math is a strong example. It is a WordPress SEO plugin that, according to its own product pages, supports 20 or more schema types, runs an SEO analysis against 30 known factors, and includes an AI visibility feature that tracks sentiment across ChatGPT and benchmarks a site against competitors. Its PRO tier is listed at €7.99 per month billed annually, excluding VAT, on the Rank Math pricing page, renewing at €8.99 per month plus taxes. If your site is on WordPress and your gap is missing structured data and badly chunked pages, that plugin will move more for you in month one than any retainer will.
Where that stops working is corroboration. No plugin can get you into the Reddit thread, the analyst directory, the comparison roundup or the practitioner post that a model is already reading. That work is relationship-driven, editorial and slow, and it is the part DerivateX is built for.
| Option | Best for | Cost | What it will not do |
|---|---|---|---|
| WordPress SEO plugin such as Rank Math | On-site structure, schema, chunking hygiene | From €7.99 per month, PRO tier, billed annually ex VAT | Build evidence or corroboration off your domain |
| AI visibility tracking tool | Teams with in-house capacity who need the measurement layer only | Varies by vendor and prompt volume | Change anything, it reports the gap weekly |
| DerivateX Diagnostic | Deciding whether the problem is entity, evidence, retrieval or corroboration | $3,500 one time, delivered in two weeks, credited in full against month one if you convert within 30 days | Execute the fixes, it tells you which ones matter |
| DerivateX Rank & Get Found | $5M to $50M ARR SaaS companies needing execution across SEO and GEO | $5,000 retainer plus $1,000 to $1,200 off-site budget, so $6,000 to $6,200 all in, 90-day pilot with no lock-in after | Guarantee a citation or a ranking, nobody can |
| DerivateX Own Your Category | Brands in contested categories where a competitor already owns the recommendation | $8,000 retainer plus $1,500 to $2,000 off-site budget, so $9,500 to $10,000 all in | Work below roughly $5M ARR, the math does not hold |
The floor at DerivateX is $5,000 per month for the retainer. Below about $5M ARR, that is usually the wrong allocation of a marketing budget, and we say so on discovery calls. The engagements where this pays back are the ones where a single closed deal covers several months of spend. REsimpli became the most cited and recommended real estate CRM for investors in ChatGPT within 90 days, and Verito moved from an average position of 40 on Google to first page while getting cited and recommended on ChatGPT and Google AI Overviews for 40 of their commercial hosting queries. Both were categories with a clear evidence gap and a buyer who asks assistants before asking sales.
Stop thinking about a ranking position and start thinking about a candidate set. Google gives a buyer ten options and lets them choose. An assistant gives them 3 to 4 and quietly discards everyone else, so the entire competition is about whether your brand survives into that short list with a description the model can defend. Publishing more pages does not change survival odds. Being resolvable, verifiable, retrievable and corroborated does, which is the whole basis on which DerivateX sequences work.
Run the grid before you run anything else: twenty-five prompts, four engines, three columns, thirty minutes. Whichever row is worst tells you which of the seven reasons above applies to you, and it is almost never the one your content calendar is built around. Everything else on this page is downstream of that diagnosis, including the tactical steps in our guide to how to rank in ChatGPT.
Because rankings and recommendations use different inputs. 80% of URLs cited by ChatGPT and Perplexity are not in Google’s top 100. A model needs a resolvable entity, verifiable evidence and corroboration from sources outside your domain before it names you. Page one on Google supplies none of those three things on its own.
Entity clarity, machine-readable proof of specific claims such as pricing and integrations, passage-level extractability, and third-party corroboration across review sites, communities and comparison content. Perplexity in particular draws 46.7% of its top sources from Reddit, so owned-domain publishing alone rarely changes shortlist position in that engine.
Lock a panel of 25 to 40 buyer prompts, run them weekly across ChatGPT, Gemini, Perplexity and Claude, and record mention, position, qualifier and cited source as four separate fields. Keep an untreated control set so model updates do not get misread as campaign results. Then match inbound leads against it.
Yes, they fail for different reasons. A mention without a citation means the model knows you but did not retrieve your page, which is a retrieval problem. A citation without a recommendation means your content was used as background while a competitor got recommended, which is an evidence problem inside the passage.
DerivateX typically sees the first movement on mention rate in 4 to 8 weeks and position changes in 6 to 12 weeks, depending on how fast third-party sources refresh. Engagements run as 90-day pilots for this reason. These are targets based on prior accounts, not guarantees of any specific outcome.
Get a free AI visibility audit from DerivateX and you will receive, within 48 hours, your current mention and citation position across ChatGPT, Gemini, Perplexity and Claude for your real buyer prompts, plus which of the seven failure modes above is your primary bottleneck: request the free AI visibility audit.
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.
