How AI Engines Decide What Your Company Does
It usually arrives secondhand. A board member or a new sales hire asks ChatGPT what your company does and forwards the answer, and the answer describes the business you were three years ago: the product line you divested, the category you left, the acquired brand listed as a competitor. Your homepage says none of that. AI engines found it anyway, because your homepage was one source among dozens.
eight25 is a Contentful Solution Partner and Contentstack Preferred Partner that builds enterprise content architectures for organizations including NetApp, Sophos, and Qlik, with structured, machine-readable content treated as part of the architecture rather than an SEO afterthought. The short answer to how AI engines decide what a company does is that they build an answer from a mixture of learned knowledge and whatever relevant sources they retrieve at query time. When those sources describe the company consistently, the engine has a much clearer signal. When the website, schema, locales, press archive, acquired brands, and third-party profiles disagree, the model has several competing versions of the company to reconcile.
AI engines answer from two memories, and they disagree about when you are
Recognition is rarely the problem. LLM Listed’s 2026 study of 250 companies across ChatGPT, Gemini, Claude and Perplexity found that AI correctly identified what a business does 92 percent of the time. But 28 percent of responses still omitted important products, services or specialist capabilities, and only 46 percent of companies received broadly consistent recommendations across all four platforms. In other words, an AI engine may know who you are while still presenting an incomplete or materially different version of the business depending on the platform and question. Being named is the easy part. Being described correctly is where enterprises lose, and a widely repeated wrong description does more damage than a quiet one.
The divergence comes from two broad sources of knowledge. The first is information encoded in the model from its training and subsequent provider-side updates: patterns learned from the data available to the model at those stages. You cannot directly edit that knowledge by changing a webpage today, so a recent repositioning may not be reflected in answers that rely primarily on the model’s existing knowledge.
The second is retrieval: information fetched from the web or other connected sources when the product searches or grounds an answer. That gives the system access to newer information, but only if the relevant pages are crawlable, retrievable and clear enough to be selected and interpreted correctly.
Google’s documentation states that AI Overviews and AI Mode may use a “query fan-out” technique, issuing multiple related searches across subtopics and data sources to build a response. A question about what you do can therefore expand into related questions about what you sell, which problems you solve, how your products fit together, who you compete with, and what independent sources say about you. Those retrieval paths can surface different URLs, domains and versions of the company story.
Not every AI answer relies on live retrieval in the same way, and different products may retrieve different sources even for similar prompts. When an answer relies more heavily on older model knowledge, or when retrieval surfaces stale or contradictory information, an earlier version of the company can reappear. That is one reason the same organization may be described accurately in one AI experience and with outdated positioning in another.
Most citations in branded AI answers may come from outside your own site
In a January 2026 analysis of 23,387 citations from 240 branded prompts across ChatGPT, Perplexity, Gemini, AI Mode and AI Overviews, Omniscient Digital found earned media made up 48 percent of citations and brand-owned content 23 percent. The split moved sharply with the question. On customer-review questions, earned sources took 82 percent. On functionality and integration questions, brand-owned content reached 50 percent, its strongest showing of any intent type.
Brand-owned content plays its strongest role when the question is about what a product does, how it works or what it integrates with. Third-party sources become more influential when the question shifts toward reputation, comparison or evaluation. “What does this company do?” can draw on both, which is why a strong owned-site description matters but cannot be treated as the only version an AI engine will encounter. A stale external source can still exert disproportionate influence when it is repeatedly retrieved or duplicated elsewhere. Seer Interactive documented how one old client review became an outsized AI signal. Pages carrying the duplicated review appeared as sources in 38 percent of 1,152 branded prompt outputs they tracked, helping sustain a recurring “high account manager turnover” narrative. After Seer published clearer first-party information and continued testing, the narrative changed across the AI responses they monitored. The lesson is not that one page can reliably “overwrite” an LLM. It is that precise, current, crawlable first-party content can give retrieval systems a stronger source when the existing record is stale or misleading.
| Surface | Who writes it | What it typically gets wrong | How you influence it |
|---|---|---|---|
| Wikipedia | Volunteer editors under conflict-of-interest rules | Can trail a repositioning badly, cites old press | Propose changes on the talk page with independent sources |
| LinkedIn, Crunchbase, G2, partner directories | You, partially | Category fields set at founding and never revisited | One named owner, updated the week positioning changes |
| Review sites and forums | Customers, former employees | Old complaints presented as current | A steady flow of recent reviews; public responses |
| Analyst and trade press | Analysts, journalists | You, in the category they cover you in | Briefings and one consistent boilerplate |
Third-party sources collectively account for a large share of the evidence AI systems cite about brands, so any serious AI visibility program has to extend beyond the company website. But the most controllable place to start is still your own content estate: the pages, fields, schema, locales and archives where your company describes itself. How to measure AI citations, a four-layer framework
The easiest contradictions to eliminate are the ones on your own site
Count the places an enterprise website describes the company: homepage, about page, footer, careers page, Organization schema, metadata, product pages, regional and language variants, acquired-brand sites, investor relations and press boilerplates. An AI system does not experience those pages as one coordinated brand narrative. It encounters individual documents and data structures that may have been written by different teams at different points in the company’s history.
That is where content architecture starts to affect AI readability. If the same entity, product, category or company description is maintained independently in dozens of templates and hundreds of pages, drift is almost guaranteed over time. If those descriptions are modeled as reusable structured content with explicit ownership, references and localization rules, the organization can change the underlying fact once and propagate it consistently across the surfaces it controls.
Take a hypothetical data company two years after divesting its backup business and acquiring a data catalog vendor:
| Surface | What it says | Last revised | Pages |
|---|---|---|---|
| Homepage hero | “Data governance platform for regulated enterprises” | 2026 | 1 |
| Press release boilerplate | “A leading provider of backup and recovery software” | 2015 to 2025 | 112 |
| Acquired brand microsite | “An independent data catalog company” | 2024 | 140 |
| German and Japanese homepages | “Data protection and backup” | 2024 | 2 |
| Organization schema description | “Backup and recovery software” | 2021 | Site-wide |
| LinkedIn company page | “Backup software” | 2019 | 1 |
The new positioning wins on one surface. Everywhere else, the company is still a backup vendor and its catalog product is a separate business.
Locale variants are a common source of drift because translations, regional approvals and publishing schedules do not always move at the same pace as the source content. That problem becomes visible at enterprise scale. In eight25’s Qlik migration from Sitecore XP to Contentful, the program covered roughly 5,000 pages across four regions and 19 languages. At that scale, keeping company positioning, product terminology and entity relationships synchronized cannot depend on editors remembering to update every instance. It becomes a content-model and governance problem. The AI visibility checks your CMS migration QA needs
A replatform either fixes this or entrenches it. Legacy pages left live, a press archive migrated wholesale, and acquired-brand content moved intact all carry the old description forward.
Decide the sentence once, then govern it like any other field
Write the canonical description first: three or four declarative sentences that state the category in the language buyers use, what the company does, who it serves and, where useful, what it does not do. Generic claims such as “leading” contribute far less entity information than concrete descriptions of products, audiences and categories. An explicit non-fit statement can also reduce ambiguity around categories or competitors the company is frequently confused with. If the business has changed materially, state the history clearly rather than assuming old descriptions will disappear from the web.
# Canonical company description: one content entry, one owner
category: "Data governance platform for regulated enterprises"
what_it_does: "Catalogs, classifies, and enforces access policy on data across cloud warehouses and on-premise stores."
serves: "Security and data teams at financial services and healthcare companies with more than 5,000 employees."
not: "Not a backup or recovery product."
history: "Divested its backup business in 2024. Acquired CatalogCo in 2024; CatalogCo is now part of the platform."
owner: "VP Product Marketing"
review_trigger: "Any acquisition, divestiture, rename, or category change"Then make it structured content, not repeated copy. The company description should exist as a governed content entry whose fields can be referenced by the footer, about page, schema layer, careers content, current press boilerplate and locale variants rather than being copied into each surface independently. Localization workflows can branch from that source entry with review and approval rules instead of creating disconnected versions. Historical press releases can retain their original text while displaying a current company description separately. Entity consistency stops depending on editors remembering every occurrence and becomes something the content architecture is designed to preserve. Schema markup and structured data, what still earns its place in 2026
None of this requires a special “AI SEO” technical layer. Google states that AI Overviews and AI Mode require no additional AI-specific files, markup or schema beyond the fundamentals used for Google Search. The architectural advantage of a well-modeled headless CMS is different: content can exist independently of presentation, with shared entities and reference fields feeding multiple pages, channels and locales from governed sources. That reduces duplication and makes consistency easier to enforce programmatically. Traditional or hybrid implementations of platforms such as AEM and Sitecore can achieve similar governance, but the more company facts are embedded directly into page templates or duplicated as page-level copy, the more operational discipline is required to keep them synchronized.
The effort is not always warranted. A single-product company with stable positioning and one locale should fix its homepage and LinkedIn page and move on. And when the misportrayal traces to reviews or press, site work will not reach it; a communications program is the larger lever there.
Measure how you are portrayed, not whether you are named
AI visibility programs often begin by measuring mentions and citations. For established brands, that is only part of the problem. The more useful second measure is whether AI describes the company accurately: category, products, audiences, exclusions, relationships and current positioning, scored against a defined canonical version.
Build the prompt set in buyer language. What does the company do. Who are its main competitors. Is it a good fit for a 2,000-person healthcare company. What is it not good for. What happened to its backup business. Score each answer for correct category, correct offering, no retired products, no acquired brand presented as a competitor, and a sensible competitor set.
Run it across models and in every language you sell in. A July 2026 variance-decomposition study of LLM brand answers found that query language alone accounted for 26.5 percent of the variance in a single response. That makes multilingual testing materially different from simply repeating the same English prompt. For an international brand, testing one assistant, one language and one phrasing provides a very narrow view of how the company is being represented.
When an answer is wrong, record which sources it cited. The citation trail names the surface to fix, and it is usually one from the inventory above.
Frequently Asked Questions
Why does ChatGPT describe my company incorrectly?
An AI assistant may answer from knowledge encoded in the model, from information retrieved at query time, or from a combination of the two. Both can draw on far more than your current homepage. If your website, press archive, locale pages, directories, reviews and third-party profiles describe the company differently, the system may encounter contradictory evidence and surface an incomplete or outdated interpretation. The more clearly and consistently the current company is described across authoritative sources, the easier that ambiguity is to reduce.
How long does it take AI engines to reflect a rebrand or repositioning?
There is no universal refresh interval. Answers that rely on live retrieval can change relatively quickly after the sources they retrieve are updated; Seer Interactive has documented stale-source and own-site corrections appearing in AI responses on a days-to-weeks timeline. Knowledge encoded in the underlying model changes on the AI provider’s schedule rather than when you update your site. Clear, current and consistent source material improves the chance that retrieval-based answers reflect the new positioning sooner, but no publisher can guarantee how quickly every AI platform will update.
Do I need llms.txt or special schema for AI engines to understand my company?
Not for Google Search. Google explicitly says that AI Overviews and AI Mode do not require new machine-readable files, AI-specific markup or special schema beyond the normal requirements of Search. Accurate Organization structured data can still provide explicit machine-readable information about the entity, particularly when it matches the visible content and official profiles, but it is not a special AI-ranking mechanism. Google has also clarified that llms.txt does not positively or negatively affect visibility in Google Search; other AI services may choose to use it, so its value remains platform-specific rather than established as a universal requirement.
Can I edit my company’s Wikipedia page to fix how AI describes us?
If you are connected to the company, directly editing the article is generally discouraged under Wikipedia’s conflict-of-interest guidance. The safer route is to disclose the connection and propose changes on the article’s talk page, supported by reliable independent sources. For a significant repositioning, earning independent coverage first gives editors stronger sourcing for an update.
What kind of agency helps fix how AI engines describe a company?
Look for an agency that can work below the copy layer. At enterprise scale, AI-readability problems often involve content models, taxonomy, structured data, localization, migrations, rendering and governance as much as messaging. eight25 approaches that work at the architecture level: defining reusable entities and governed fields in the CMS, mapping where those entities appear across pages and locales, and making sure the implementation exposes consistent, crawlable information in the rendered site. That is where headless platforms are especially useful: the same structured entity can serve the website, schema, regional experiences and other channels without maintaining a separate copy of the underlying fact in each one. Earned-media and reputation signals still sit outside the CMS and require a communications strategy of their own.
Before the next repositioning, count how many versions of you are live
The worst time to discover an AI-description problem is after a repositioning, acquisition or platform launch, when the new story is live but older versions remain crawlable throughout the estate. The better sequence is to inventory the surfaces that describe the company, define the canonical entities and relationships, and encode them into the content model before launch. eight25 uses that inventory to identify where duplicate descriptions, legacy pages, locale variants and acquired-brand content can create contradictory machine-readable signals. If a rebrand, acquisition or replatform is already on the roadmap, that is the natural point to make AI readability part of the architecture rather than a cleanup project afterward.