Case study · Custom AI Systems
Making a company legible to AI search, not just Google
When someone asks an AI assistant about a company, the answer is assembled from whatever the model can retrieve. The work is structuring a company's information so generative engines retrieve, understand, and attribute it correctly, rather than describing it from a stale or scraped guess.
- 01Machine-readable facts
- 02llms.txt summary
- 03Entity authority
- 04Retrieval-friendly structure
- 05AI-crawler access
- Client
- a B2B technology company
- Problem
- AI search systems answer about a company from whatever they can scrape, often wrong or outdated.
- Approach
- Machine-readable content, an llms.txt summary, entity authority, source attribution, retrieval-friendly structure.
The problem
When someone asks an AI assistant about a company, the answer is assembled from whatever the model can retrieve. If the company's own information is unstructured, inconsistent across pages, or absent from machine-readable surfaces, the engine fills the gap with stale or incorrect detail, and there is no clean source for it to attribute. The company does not control how it is described at the moment it matters most.
Traditional SEO does not solve this. Generative engines retrieve and synthesize rather than list links, so the work is making a company's facts and expertise retrievable, consistent, and attributable across the surfaces an engine actually reads.
The reality behind the problem
Ranking on Google does not carry over. A page can rank first in a results list and still be invisible to an engine that answers by retrieving passages and composing them, because that engine is not reading the page the way a search crawler ranks it. It is looking for a clean, consistent, machine-readable statement of what the company is and does, and most sites do not provide one.
The deeper problem is consistency. When the founder's name, the offering, or the company's category differ from page to page, or between the site and the third-party profiles an engine trusts, the engine hedges. It merges the company with a similarly named one, describes it vaguely, or repeats an outdated fact. No amount of keyword work fixes a source that contradicts itself.
What was assumed
If we rank on Google, AI search will describe us correctly.
What Stallwart asked
What can an LLM actually retrieve about us, and is it consistent and current?
The desired outcome
The target is not a ranking. It is control over the source a generative engine reads, so the description it produces is accurate and attributable.
- A canonical, machine-readable statement of what the company is, does, and serves.
- The same core facts everywhere: across pages, structured data, and summary surfaces.
- Passages a retrieval system can lift cleanly and attribute back to the company.
- One well-described entity, disambiguated from similar names, not a fuzzy one.
- Explicit access for the AI crawlers that matter, instead of accidental blocking.
- Machine surfaces that stay in sync with the live site, so they never become a new source of wrong answers.
The system we designed
Stallwart makes a company's information retrievable and consistent for generative engines. Core facts, offerings, and expertise are stated in a structured, machine-readable form and kept consistent across the site, so a model retrieves the same answer wherever it looks. A plain-text summary surface, an llms.txt, gives AI systems a canonical, quotable description rather than leaving them to infer one.
Content is organized so a retrieval system can find a clean, attributable passage for a given question, and entity authority is strengthened so the company maps to a single, well-described entity. Stallwart runs exactly this on its own site: an llms.txt and llms-full.txt, an explicit allowlist for AI crawlers such as GPTBot, ClaudeBot, and PerplexityBot, and Organization structured data that states the entity plainly. All of it is publicly verifiable.
How it works
- 01
Give AI a canonical source, not a guess
A machine-readable summary and consistent structured facts give a generative engine one clear thing to retrieve, instead of synthesizing a description from scattered or stale fragments.
- 02
Make passages attributable
Content is structured so a retrieval system can lift a clean passage and point back to its source. Attribution is how a generative answer can credit the company rather than paraphrase it anonymously.
- 03
Strengthen the entity, control the access
Clear entity signals map the company to a single well-described entity, and an explicit crawler allowlist lets the answer engines that matter read the site, the approach Stallwart uses on its own site.
How it works
From scraped guesswork to a controlled, retrievable source.
- 01
Machine-readable facts
Structured, consistent across pages
- 02
llms.txt summary
A canonical description for AI systems
- 03
Entity authority
One well-described entity, disambiguated
- 04
Retrieval-friendly structure
Clean, attributable passages
- 05
AI-crawler access
Explicit allowlist for answer engines
Each layer gives generative engines a cleaner source to retrieve and attribute.
Architecture
What sits underneath.
Underneath the approach are four surfaces an engine reads, each kept consistent with the others so there is one story to retrieve.
Machine-readable content and structured data
Organization and related schema state what the company is, does, where it operates, and who founded it, in a form an engine parses directly rather than inferring from prose.
Canonical summary surface
An llms.txt and a fuller llms-full.txt give AI systems a plain-text, quotable description and the key facts, so the engine retrieves a source the company controls instead of composing one from fragments.
Entity and disambiguation signals
sameAs links tie the site to its authoritative profiles, and explicit disambiguation keeps the company from being merged with a similar name, so the entity graph resolves to one clear entity.
Crawler access and consistency checks
An explicit allowlist lets the answer engines that matter read the site, and the core facts are verified identical across pages and summary surfaces, because a contradiction is what makes an engine hedge.
The hard parts
Where the engineering judgment was.
Treat llms.txt and structured data as first-class
Generative engines retrieve from machine-readable surfaces. A canonical summary and consistent schema are the most direct way to influence how a company is described.
The trade-off
Maintenance discipline: the machine-readable surfaces must stay in sync with the live site, or they become a new source of wrong answers.
Optimize for retrieval and attribution, not inclusion promises
No one controls whether an LLM includes a company in an answer. The controllable part is being the cleanest, most consistent, most attributable source on the topic.
The trade-off
Honest framing: this improves the odds and the accuracy of being represented, it does not guarantee a citation.
Spell out what the company is not
When a name resembles a common word or another business, leaving disambiguation implicit invites an engine to merge or confuse entities. Stating what the company is and is not resolves it.
The trade-off
Writing what you are not reads oddly in marketing copy, but it is exactly what machine understanding needs.
Reliability & guardrails
How it avoids the wrong call.
The guardrails are about keeping the source honest, so the engine never has a contradiction to trip on.
One set of facts, everywhere
The founder, the offering, the category, and the location are stated identically across pages, structured data, and the summary surfaces, because a single contradiction is enough to make an engine hedge.
Attribution integrity
Passages are written to be liftable and traceable to the company, so a generative answer can credit the source rather than paraphrase it anonymously or attribute it elsewhere.
Disambiguation held explicitly
The entity is distinguished from similar names in structured data and in the llms.txt, so an engine resolves to the right company instead of merging it with another.
Honest scope
The system improves accuracy and the chance of being represented. It makes no claim to guarantee inclusion or citation, which remains the engine's call.
Productionization
What turns it from a demo into software.
What keeps the machine-readable surfaces trustworthy over time, rather than correct once and stale after.
Summaries generated from source
The llms.txt and llms-full.txt are generated from the site's own data, so they track the live content instead of drifting into a separate, outdated copy.
Structured-data validation
Schema is validated as part of the build, so a broken or incomplete entity definition is caught before it ships rather than after an engine reads it wrong.
Crawler-access checks
The AI-crawler allowlist and robots rules are kept explicit and reviewed, so the engines that matter are not accidentally blocked and access stays intentional.
Consistency monitoring
Core facts are re-checked across pages and summary surfaces when content changes, so the source stays internally consistent as the site grows.
What changed
The change is in who controls the description. Instead of an engine composing an answer from scattered, stale fragments, it retrieves a canonical, consistent, attributable source the company owns, and resolves to one clear entity rather than a fuzzy guess.
This is the posture Stallwart holds on its own site, which anyone can verify by reading its llms.txt and llms-full.txt and checking its AI-crawler allowlist. It improves the accuracy and the odds of being represented well. It does not, and does not claim to, guarantee a citation.
- Generative engines retrieve a source the company controls, not a scraped guess.
- The company's facts read the same across pages, structured data, and summaries.
- The company resolves to one clear entity instead of being merged with another.
- The answer engines that matter can read the site, by explicit allowlist.
Ownership & handover
What you receive, and keep.
At handover, the machine-readable surfaces and the system that keeps them current are yours.
- The llms.txt and llms-full.txt, and the generation that keeps them in sync with the site.
- The Organization and entity structured-data templates.
- The AI-crawler allowlist and robots configuration.
- Runbooks for updating facts, regenerating summaries, and validating schema.
- Documentation of the entity model and the consistency checks.
What we learned
The insight that generalizes: GEO is a source-control problem, not a ranking one. You cannot control the answer a generative engine gives, only how clean, consistent, and attributable the source it reads is. Make the company legible to a machine, keep it legible as the site changes, and the accuracy of how it is described follows.
This work connects to
- Custom AI Systems
- Generative engine optimization
- AI search visibility
- Entity authority
- llms.txt and machine-readable content
Frequently asked
What is generative engine optimization (GEO)?
GEO is making a company's information retrievable, consistent, and attributable for generative search systems and LLMs, so that when an AI assistant answers about the company it uses accurate, controlled sources. It relies on machine-readable content, a canonical summary such as llms.txt, strong entity signals, and retrieval-friendly structure.
How is GEO different from SEO and AEO?
SEO targets rankings in a results list and AEO targets extractable answers like snippets. GEO targets how generative engines retrieve and synthesize information about an entity, which puts the emphasis on machine-readable facts, consistency across sources, entity authority, and attribution.
Can you guarantee an AI like ChatGPT or Perplexity will cite us?
No. No optimization guarantees inclusion or citation in a generative answer. GEO improves accuracy and the chance of being represented by making a company the cleanest, most consistent, most attributable source on its topic. Selection remains with the engine.
Does Stallwart do this on its own site?
Yes, and it is publicly verifiable. Stallwart publishes an llms.txt and llms-full.txt, maintains an explicit allowlist for AI crawlers such as GPTBot, ClaudeBot, and PerplexityBot, and states its entity in Organization structured data. The approach it applies for clients is the one it runs on itself.
What is an llms.txt file and why does it matter for GEO?
An llms.txt is a plain-text summary of a site written for AI systems, giving them a canonical, quotable description and the key facts instead of leaving them to infer one from scattered pages. It matters for GEO because it gives a generative engine a clean source the company controls to retrieve and attribute.
Does AI search describe your company from a guess?
We make your facts and expertise retrievable and consistent, so generative engines have a clean source to work from.
Last updated: October 13, 2026