Why you're not cited in AI search: the three failure modes a prompt re-run tells apart

Questions this guide answers

• Why am I not cited in AI search?
• How many times should I re-run my AI prompts before believing a change?
• Why does my brand show up on Gemini but not ChatGPT?
• How do you diagnose an AI citation gap?

Direct answer: a re-run answers one question

If you are not cited in AI search, the useful first move is not to write more content. It is to find out which of three things is actually broken, because each one has a different fix and they do not overlap.

A re-run of your buyer-prompt set - the same prompts, asked again across the same engines - exists to sort each result into exactly one of three mutually exclusive failure modes. Mode 1 is Absent: the answer never names you, so the problem is that no source the engine trusts corroborates you on that topic. Mode 2 is Mis-described: the answer names you but describes you flatly, with stale or thin framing, which traces back to a specific source feeding weak phrasing. Mode 3 is Out-competed: the answer names a competitor as the primary recommendation instead of you, which means they out-corroborate you in the sources that engine reads. The triage table below maps each mode to what you see in the answer, the root cause, and the lever that fixes it.

Re-running is a diagnostic, not a scoreboard

Most teams treat a prompt check the way they treat a rank tracker: a single number that goes up or down. That framing wastes the most valuable thing a re-run produces, which is not a score but a classification.

SolCrys runs a four-step loop: measure, then diagnose, then execute, then verify. Measure captures what the engines actually said. Diagnose - the step this article is about - reads that capture and assigns each weak result to one of the three failure modes. Execute ships the fix the mode calls for, under human approval. Verify re-runs the same prompt set to confirm the fix moved the answer and was not just sampling noise.

The reason the three modes matter is that they are mutually exclusive. A result is Absent, or Mis-described, or Out-competed - not two at once on the same prompt. That exclusivity is what makes the diagnosis actionable: once you know the mode, you know the lever, and you stop spending effort on the wrong one. Writing a better homepage does nothing for an Absent result, and earning more third-party mentions does nothing for a result that is already named but described badly.

These three modes are a routing layer, not a competing taxonomy. They collapse the finer-grained five answer-gap types into the smallest set of buckets that each point to a different first action.

To make each mode concrete, picture a single illustrative vendor running this diagnosis: a mid-market cloud data-warehouse company, well-known to a handful of data teams but not a category leader, tracking a buyer-prompt set across the five engines. The walk-throughs below are a teaching example, not measured results - but they show exactly what each mode looks like when it lands on a real prompt.

Mode 1 - Absent: nobody trusted enough corroborates you

The clearest signal in the data is absence. Picture our illustrative mid-market warehouse vendor running the unbranded discovery prompts a real buyer would ask - "best data warehouse for real-time analytics", "best cloud data warehouse for a small data team", "cheapest cloud data warehouse for a startup". Across runs, on those category-defining prompts, the vendor is simply never named. The answers come back full of the obvious leaders and the vendor does not appear at all.

Now watch what flips when the prompt already contains the brand. A branded check like "is [vendor] a legitimate cloud data warehouse?" comes back present on essentially every run - the engines will happily confirm you exist when the question hands them your name. But the unbranded discovery prompts, the ones that actually drive a buyer's shortlist, stay at or near zero presence the whole way through.

That shape - near-total presence on your own name, near-total absence on the category - is the signature of a missing-source problem, not a content problem. The engine has nothing to draw on because no source it trusts mentions the vendor in that context. The fix is third-party corroboration: earn genuine, value-first mentions in the kinds of sources the engine already cites - the community, editorial, and comparison content that no single brand owns a large share of - so there is something to retrieve when the buyer prompt fires.

Confirm the mode before acting on it: a result that reads as Absent on one run but flips to present on the next is sampling noise, not a stable gap, which is why the re-run discipline below comes before any fix. To route this fix, start with how to build a source-layer strategy and the breakdown of owned, earned, and community sources for AI. For how much of the citation pool your own domain actually holds in a real account, see the visibility measurement methodology rather than a number that moves week to week.

Mode 2 - Mis-described: a source is feeding flat phrasing

The second mode is harder to spot because the brand is present. The answer names you - it just describes you in a flat, neutral, or thin way that does not move a buyer. This is a description-quality gap, and it is almost always per-engine, because different engines read different sources.

Stay with the illustrative warehouse vendor, but switch to a comparison prompt where it does get named - something like "Snowflake vs [vendor] for streaming". In the same week, the framing splits cleanly by engine. On one engine the vendor comes back positively - described as a credible, capable choice for streaming workloads. On another engine, the same week, the same comparison, the vendor is named but described flatly: present, neutral, thin, nothing that would move a buyer off the leader. Read the direction, not a precise average: positive on one engine, flat on another, with no change to the brand in between.

Two clarifications matter. First, this is a measure of framing polarity and provenance - description quality - not a claim that a specific wrong or stale fact was caught in the answer string. The measurement exposes polarity, not the literal sentence. Second, because the gap is per-engine, the fix is per-source: a flat description on one engine traces to whatever source that engine pulled, and you fix the phrasing at that source, not on your homepage, which the engine may never have read.

The wrong move here is to rewrite the homepage and re-run, because the homepage was never the input. The right move is to identify the source feeding the flat framing for that engine and improve the description there. To route this fix, work through how AI describes your brand.

Mode 3 - Out-competed: a competitor owns the slot

The third mode is the most uncomfortable: you are present, the description is fine, but a competitor is named as the primary recommendation and you are not. This is not a content-quality problem on your page. It means the competitor out-corroborates you in the sources that engine reads.

Back to the illustrative vendor, now on a prompt like "best data warehouse for real-time analytics". A category leader - Snowflake, say, or Databricks - is named as the primary recommendation in the large majority of answers, while the mid-market vendor, when it appears at all, sits near the bottom of the share of voice. The gap is large, and it is not random - it is built on coverage. The primary-rate itself drifts run to run, so the right read is the shape: leader on top, vendor near the floor, consistently across the set.

The diagnostic value of this mode is that the leader's citation map is your to-do list. Reading where the leader gets cited shows you the exact sources you are missing - the comparison roundups, the editorial explainers, the community threads that named them and not you. That is the literal map of where they appear and you do not, and it tells you which doors to knock on. Note that no single source owns a large share of those citations; the cited layer is community, editorial, and comparison content, so the work is breadth of corroboration, not one big placement.

To route this fix, start with how to track where competitors get cited in AI and run a citation gap audit to turn their coverage map into your own source-acquisition list. If a buyer prompt is actively returning a rival as the answer, the ChatGPT-recommends-a-competitor self-audit walks the same diagnosis end to end.

The discipline half: re-run enough to know a change is real

A diagnosis is only as trustworthy as the number of runs behind it. One run is one draw from a distribution, and AI answers are sampled, so the same prompt set will return different results on different days even when nothing about your brand or the web has changed.

Picture the warehouse vendor running its identical prompt set on several consecutive days with nothing shipped in between. The overall mention rate still wanders day to day - up one morning, down the next, with no real trend underneath. That movement is the engine sampling, not progress or decline. If you had acted on the highest day as if it were a win, you would have been chasing a draw.

This is why diagnosis precedes any fix, and why verification re-runs rather than checking once. A single-run jump after you ship a fix proves nothing on its own; you confirm a change is real only when the prompt set moves and stays moved across enough runs to separate signal from sampling. For how many runs that takes, see how many runs until your AI visibility is trustworthy, and for why the underlying score wanders even on a stable brand, see why your AI visibility score moves.

From diagnosis to fix: the routing table

The whole point of sorting a result into one of three modes is that each mode hands you a different lever and a different SolCrys surface to execute it. The table below is the hub: find your mode, take the lever, ship the fix under human approval.

This is the Diagnose step. Execute and Verify close the loop - and Execute stays governed and human-approved, with no guaranteed citation or lift attached to any single fix. A re-run tells you what is broken and re-running again tells you whether the fix held; it does not promise that any one action will land you in the answer.

Sources

• SolCrys citation, visibility, and per-prompt measurement, workspace solcysai-aeo, 7-day window ending 2026-06-22, five engines (ChatGPT, Perplexity, Google AI Overviews, Gemini, Claude)

FAQ

Related guides