Integrated search is really a publishing decision

In September 2025, Gartner published a Market Guide for Enterprise AI Search. The useful line, as summarised by GoSearch in April, is a vocabulary one: enterprise AI search demands "APT" content, meaning Accurate, Pertinent and Trusted, and is significantly hindered by "ROT", meaning Redundant, Obsolete and Trivial material. Gartner's recommendation, per the same summary, is to combine embedded in-app search with an enterprise-wide platform, and to treat enterprise search as the foundation that AI assistants and agents sit on.

That reframing matters more than the acronyms. Integrated search sounds like a connector problem: wire the model into Confluence, Drive, Notion, the CRM, the ticketing system, done. In practice, the moment you point a generative assistant at an unfiltered corpus, you have made a publishing decision. Every answer the assistant produces is an implicit claim about which document, which version and which author to trust.

Most teams I speak to have not made that decision. They have made the opposite one by omission. Everything the company has ever written lives in the index. The quarterly plan from 2022 sits next to the current one. Three versions of the pricing page coexist. A deprecated API doc has the same weight as the live one. The assistant does what assistants do: it blends. The output is fluent, plausible and often confidently wrong on the small facts that actually matter in a customer conversation or an investor update.

The mechanism is boring, which is why it is underrated. Retrieval-augmented systems rank candidate passages, then let the model paraphrase from the top matches. If "top" is defined by vector similarity alone, redundancy wins: five near-duplicate pages outrank one canonical source. Obsolete pages often look more relevant than current ones because they were written when the topic was new and keyword-dense. The system is doing exactly what it was built to do, faithfully reflecting a corpus nobody curated for this use.

So the first question before buying an integrated search layer is not which vendor. It is narrower, and awkward: for each category of answer we want the assistant to produce, which single document is the one it should quote? If the sales team asks "what is our current enterprise pricing?", there is a right answer and it lives somewhere. If marketing asks "what did we say about data residency last quarter?", there is one post of record. If product asks "what is the agreed definition of an active user?", there is a dashboard spec. Name those sources. Mark the rest as reference, archive or ROT.

APT is a decision about status, not a quality score for individual documents. A document becomes trusted because an owner publishes it as trusted and keeps it current. ROT is the state of any file nobody has agreed to maintain, regardless of age. An integrated search tool cannot infer either. It can only enforce what a human has already decided.

The practical consequences for a small technology team are specific. First, the project plan for integrated search needs an owner for each source category, not just an engineer for each connector. Second, the index needs a visible exclusion list, reviewed quarterly, because what you keep out of the corpus matters as much as what you put in. Third, the assistant's answers need provenance on screen: the user sees which document was quoted and when it was last updated, and can challenge it. Without that, you cannot tell whether a wrong answer is a model problem or a corpus problem, and you will spend months blaming the wrong layer.

There is a second-order effect worth naming. Once a team draws the APT line internally, the same discipline shows up in what it publishes externally. The claims in a blog post, a sales deck or a LinkedIn update start to trace back to a specific internal source of record, with a date and an owner. That is what distinguishes thought leadership from content: a reader, or a model summarising the piece later, can follow the claim back to something a human stands behind.

This is the practical reason we built Matrex Terminal around a curated set of canonical sources rather than a general index. Content infrastructure for thought leadership is the layer that decides which internal facts, positions and numbers are allowed to appear in a published argument. The same APT and ROT choice Gartner describes for enterprise AI search applies, one step downstream, to anything a company says in public, and a better editor cannot substitute for it.

If I were running the integrated-search evaluation again today, I would start the first workshop without any vendors in the room. I would list the twenty questions the business most wants an assistant to answer well. For each, I would name the single source of truth and the person who owns it. The ones with no clear source or no clear owner are the real project. The search layer is the easy part.

Source: GoSearch, summarising Gartner's Market Guide for Enterprise AI Search (September 2025). https://www.gosearch.ai/blog/gartner-market-guide-enterprise-ai-search-2026/