Retrieval and grounding audit
We test your priority pages against a live prompt set and record whether they are retrieved at all, and if so, whether they get grounded into the final answer or dropped.
AEO service
AI search visibility means being the passage the model picks
Ranking well and getting cited are two different outcomes decided by two different systems. We work the second one directly, from retrieval through grounding to the sentence that actually gets quoted.
AI search visibility is the discipline of making your content eligible to be pulled into an assistant's answer and named as the source. It is not a synonym for search rankings. A page can hold the number one position on Google and still be invisible inside ChatGPT, Perplexity, or an AI Overview, because those systems are not choosing between pages. They are choosing between passages, and they are applying a different set of tests to make that choice.
The mechanism behind an assistant's answer usually runs in stages. A query gets expanded into several related sub-queries, a process often called query fan-out, so the system can gather evidence from more than one angle before it writes anything. Candidate documents get retrieved for each sub-query, broken into chunks, and scored for how well each chunk answers the specific question in isolation. The model then grounds its draft answer in a handful of the highest-scoring chunks and decides which of those chunks earn a visible citation.
Each of those stages has a failure mode that has nothing to do with authority or rank. A page can be chunked badly, splitting the actual answer across two passages so neither one scores well alone. A page can answer the question but bury it under three paragraphs of preamble the chunker never reaches. A brand can be the correct answer but appear in only one source, while three competitors are corroborated across five, and corroboration frequently outweighs a single strong source.
We treat visibility as an engineering problem with content as the interface. That means auditing how your existing pages actually get chunked and scored, not guessing from best practice checklists, and then rebuilding the specific passages that should be winning a given question but currently are not.
What it is
The mechanics of retrieval and citation
Retrieval is the step where a system pulls in candidate content before it writes anything. Some assistants retrieve live from the web, others draw on a cached index refreshed on a cycle, and some blend both depending on how time-sensitive the query looks. Understanding which mode applies to a given prompt matters, because a page published yesterday is invisible to a system running on a stale index no matter how well it is written.
Grounding is what happens after retrieval. The model does not simply summarize the top result. It weighs multiple retrieved chunks against each other, checks whether they agree, and drafts an answer anchored to the passages it trusts most. A chunk that contradicts the consensus of the others tends to get discarded even if it is correct, which is one reason a lone, unsupported claim underperforms a claim repeated with the same numbers across a few independent sources.
Chunking and passage selection decide the unit that actually competes. Most systems split a page into segments of a few hundred words before scoring them, so a page's overall quality does not matter if the one segment covering the target question is diluted by unrelated content around it. We write to the chunk: one clear question, one complete answer, minimal dependency on the paragraphs before it.
Corroboration and recency round out the picture. A claim is more likely to be surfaced and cited when comparable phrasing of the same fact shows up on more than one independently operated domain, and a claim tied to a visible date is more likely to be trusted for anything where currency matters, from pricing to product specs to local information. Neither factor is something a single page can fix alone, which is why visibility work always looks outward from the page toward the rest of the web that talks about you.
Fit
Who this is for, and who it is not for
We would rather say no early than sell a program that cannot work.
Good fit
- Brands that already rank reasonably well but see zero citation share when they run their own prompts.
- Categories where buyers ask genuinely comparative or how-to questions before purchasing.
- Teams willing to restructure existing pages rather than only publish new ones.
Not a fit
- Sites with no indexable content for an assistant to retrieve in the first place.
- Categories where nobody phrases a question before buying, so there is no prompt to win.
- Anyone expecting a guaranteed citation rate. This surface does not offer that kind of certainty yet.
Deliverables
What you get
Diagnosis of how your pages are actually being retrieved, then rewritten passages built to win.
Chunk-level rewrite
The specific passage answering each target question rebuilt to stand alone, so it scores well as an isolated chunk rather than depending on surrounding context.
Corroboration map
A list of independent sources that already state or could reasonably state the same facts about you, prioritized by which gaps are most likely to cost you a citation.
Recency signal audit
Visible publish and update dates, changelog patterns, and freshness cues reviewed against what time-sensitive queries in your category actually need.
Query fan-out coverage
The related sub-questions a real buyer prompt expands into, mapped against your existing content so gaps in the fan-out get identified before a competitor fills them.
Prioritized fix list
Every finding ranked by expected effect on citation share, not treated as an equal-weight checklist, so the highest-leverage passages get rewritten first.
How we run it
How we run it
We start from what actually happens when an assistant answers your prompts today, not from a generic audit.
Step 1: Build the real prompt set
We collect the actual questions your buyers ask, including the follow-up and comparison questions a query fan-out is likely to generate.
Step 2: Run and record
The prompt set gets run across the major assistants, and we log whether you are retrieved, grounded, cited, and in what position and tone.
Step 3: Diagnose the failure point
For each miss, we identify whether the issue is retrieval, chunking, grounding, corroboration, or recency, because the fix is different for each.
Step 4: Rewrite at the passage level
Priority passages get rebuilt to answer completely in isolation, and we close corroboration gaps where a fact only lives on one domain.
Step 5: Re-run and adjust
The prompt set gets re-run on a schedule so we can tell whether a fix actually changed citation behavior or just changed the page.
Why can a page rank first on Google and never get cited by ChatGPT?
Ranking measures a page against a query. Citation measures a chunk of that page against a specific question, checked for standalone clarity and corroboration elsewhere. A page can rank well overall while the exact passage an assistant needs is diluted, buried, or unsupported by any other source.
The gap shows up most often on pages where the answer is real but spread across several paragraphs, or where the brand is the only source stating a given fact.
- Chunking can split an answer across two passages, so neither scores well alone.
- A single unsupported claim often loses to a corroborated claim from a weaker page.
- Stale or undated content underperforms on time-sensitive queries even when it ranks well.
What is query fan-out and why does it matter for visibility?
Query fan-out is when an assistant expands one user question into several related sub-queries before retrieving evidence. A page that only answers the literal question misses the sub-queries entirely, while a page covering the surrounding angles has more chances to be retrieved and cited.
We map the fan-out for each priority prompt so content coverage is planned around the full set of sub-questions, not just the headline one.
Where this connects
Where this connects
Visibility work depends on structure, access, and measurement that live in the rest of the AEO pillar.
None of this matters if the assistant's crawler cannot reach the page in the first place, which is why we start every engagement inside the AEO pillar with an access and rendering check before touching a single sentence.
A machine-readable summary of your priority pages gives assistants a shortcut into your best content, which is exactly what llms.txt implementation is built to provide.
Structured data cannot replace a well-written passage, but it does confirm the machine's reading of it, which is the job of schema markup layered on top of the rewritten content.
Corroboration also depends on the model resolving mentions of you across the web as one consistent entity, which is handled under entity optimization rather than inside any single page.
None of this is worth doing without a way to tell if it worked, which is why the prompt-run results feed directly into AI citation tracking on a recurring schedule.
Questions
AI search visibility questions we get asked
See whether the assistants are actually citing you today
We run your real prompt set before we recommend a single change.