Cognitive Search Platforms

Enterprise search spent twenty years as one of the least glamorous line items in IT. Everyone had it, nobody liked it, and the running joke was that finding anything internally was harder than finding it on the open web.

Then retrieval-augmented generation arrived and the search index stopped being a convenience. It became the thing that determines whether your AI tells employees the truth.

That is the repositioning underneath this category. Cognitive search is no longer primarily about helping a person locate a document. It is the layer that decides what an AI system knows about your organisation before it answers, which makes retrieval quality a correctness problem rather than a usability one.

What the platforms actually do

Forrester's definition, from its landscape research, describes cognitive search platforms as knowledge and information retrieval platforms covering the entire search lifecycle: data ingestion and indexing, querying and search, and insights delivery.

Three stages, and the difficulty is not evenly distributed.

Ingestion and indexing means connecting to every system holding content. File shares, SharePoint, Confluence, ticketing systems, CRM, email, the wiki nobody maintains, the PDF repository from an acquisition. Each connector has to authenticate, crawl, extract text from whatever format it finds, and keep up with changes.

Querying and search is where the cognitive part sits. Where traditional search matched keywords and ranked by frequency, these platforms apply natural language processing, semantic understanding, vector embeddings, knowledge graphs, and machine learning relevance models to work out what someone meant rather than what they typed.

Insights delivery is what happens with the results. A ranked list, a synthesised answer with citations, a summary, or increasingly a response inside another application entirely.

The interesting shift is that the third stage has quietly become the second stage's customer. Increasingly the thing consuming search results is not a person scanning a list but a language model assembling an answer.

Why RAG raised the stakes

A language model asked about your organisation knows nothing about it. Retrieval-augmented generation solves that by searching your content, pulling back the relevant passages, and giving the model that material to answer from.

Which means every property of your search system becomes a property of your AI system.

If retrieval returns the wrong documents, the model answers confidently from the wrong documents. If it returns an outdated policy, the model states outdated policy as current fact. If it retrieves nothing useful, the model fills the gap from its training data, which is the specific mechanism behind a large share of enterprise AI hallucinations.

The failure is invisible in a way that traditional search failure was not. A person who receives a bad list of search results knows the search failed and tries again. A person who receives a fluent, well-structured, cited answer built on the wrong retrieval has no signal that anything went wrong.

So the unglamorous middle layer became the thing that determines whether the impressive layer on top can be trusted. Kore.ai's framing of cognitive search as the brain behind agentic AI is vendor language, but the underlying claim is fair: an agent that acts on retrieved information is only as reliable as the retrieval.

The permission problem got existential

Enterprise search always had to respect access control. A search result must not surface a document the user cannot open, and every serious platform has done document-level security trimming for years.

Generative answers break that model in a specific and dangerous way.

Traditional search either shows you a document or it does not. Generative retrieval can summarise a document without showing it to you. If the permission check happens at display time rather than at retrieval time, a user can receive an accurate summary of a compensation review, an acquisition memo, or a disciplinary file they have no right to see, without the system ever having shown them a link.

The information leaks through the answer rather than through the result.

This is why security scored so prominently in the current evaluation, and it is the single most important thing to interrogate in a demonstration. The question is not whether the platform supports permissions. Every vendor says yes. The question is whether permission filtering happens before content reaches the model, whether permissions are evaluated live or from a cached copy that may be days stale, and what happens when someone's access is revoked.

A cached permission model that refreshes nightly means a departing employee's access changes are a day late in a system that can now summarise anything it retrieves.

Inside The Forrester Wave: Cognitive Search Platforms, Q4 2025

Published on 3 October 2025 and authored by Rowan Curran, the evaluation scored fourteen providers against twenty one criteria, sorting them into Leaders, Strong Performers, and Contenders.

Mindbreeze placed as a Leader with the top ranking in the current offering category among all fourteen vendors, positioned by Forrester as a fit for enterprises wanting a platform that is both highly secure and highly flexible. That combination of top current offering plus a security-first characterisation is consistent with where the category's difficulty has moved.

Kore.ai placed as a Leader with the highest score in the strategy category and the maximum score in eleven criteria. Forrester noted that it had capitalised on the enterprise rush to deploy conversational AI, which is a precise description of how a company from the chatbot world ended up leading a search evaluation.

Elastic placed as a Leader on the basis of Elasticsearch and the wider platform it anchors alongside Logstash and Kibana, supporting a broad range of search-driven use cases with observability and compliance strength. Forrester positioned it for organisations wanting to build and customise their own search experience rather than buy a finished one, which is a meaningfully different purchase from the other two.

Moveworks placed as a Strong Performer in its first participation in this evaluation, arriving from employee support automation rather than from search.

The remaining placements are not fully public.

What changed since 2023

The predecessor, The Forrester Wave: Cognitive Search Platforms, Q4 2023, also evaluated fourteen vendors but against twenty seven criteria, scoring current offering, strategy, and market presence. Sinequa placed as a Leader with the highest current offering score in that edition.

Two years later the vendor count is identical and the criteria count has dropped by six.

That contraction is worth reading. Criteria fall out of an evaluation when they stop discriminating between vendors, which is what happens as capabilities become table stakes. Vector search, semantic ranking, and natural language querying were differentiators in 2023 and are baseline now, so the evaluation narrowed to where the differences actually are.

The composition of the leadership also moved. The top current offering score passed from a specialist enterprise search vendor to a different one, and a conversational AI company took the strategy lead. That last part is the structural change. The boundary between enterprise search and enterprise assistant has collapsed, and vendors from either side now compete for the same budget.

The part nobody demonstrates

Connectors decide whether any of this works, and they are boring enough that they rarely feature in an evaluation.

Every platform ships an impressive connector list. What the list does not tell you is the fidelity of each one. Whether it captures permissions from the source system or only content. Whether it handles incremental updates or re-crawls everything. Whether it extracts text from a scanned PDF or indexes an empty document. Whether it understands the nested folder permissions in your particular SharePoint configuration, which is not the same as supporting SharePoint.

The pattern is consistent across implementations. The two or three systems holding your most valuable content are usually the ones with the weakest connector coverage, because they are old, customised, or niche. Those are also the systems people most need searched.

There is a related question about what happens to content quality. A retrieval system indexing five years of contradictory policy documents will faithfully retrieve contradictory policy documents, and a generative layer on top will confidently synthesise an answer from whichever it found. The platform cannot tell you which version is current unless something in your metadata does.

Which means content governance, the least fashionable discipline in this entire area, has become a prerequisite for AI reliability rather than a documentation hygiene issue.

Chunking is a design decision

One technical detail worth understanding because it materially affects output quality and vendors handle it differently.

Retrieval systems do not return whole documents to a model. They return passages, which means every document gets split into chunks at indexing time. How that split happens determines what the model sees.

Chunk too small and you sever context, returning a paragraph that references a condition stated three paragraphs earlier. Chunk too large and you dilute relevance, burying the answer inside material that pushes the model toward the wrong emphasis. Split a table across chunks and the numbers arrive without their headers.

Good platforms chunk with structural awareness, respecting document hierarchy, keeping tables intact, and preserving the surrounding context a passage needs to make sense. Weaker ones split on a fixed character count.

You cannot see this in a demonstration and it is worth asking about explicitly, because it is one of the largest determinants of answer quality and one of the least discussed.

What to test

Bring your worst content. Not the clean policy library. The shared drive with fifteen years of accumulation, duplicate files, and documents whose owner left in 2019. That is what the system will actually index.

Test permissions adversarially. Create a test account with restricted access, then ask questions whose answers live in documents that account cannot open. Try several phrasings. If any answer contains information from those documents, whether summarised, paraphrased, or cited, you have found a leak that would be a serious incident in production.

Ask the same question five ways. Retrieval quality shows up as consistency. If phrasing changes the answer materially, the semantic layer is weaker than the demo suggested.

Ask about the documents you did not get. A good system can tell you what it searched and why it selected what it selected. A system that only shows you the answer is asking for more trust than it has earned.

Check connector fidelity against your specific configuration, not the vendor's supported list. Ask what a connector captures for permissions, how updates propagate, and how long a change in the source system takes to appear in results.

And decide whether you are buying a product or a platform. The Leaders in the current evaluation include finished experiences and infrastructure you build on, and Forrester positions them differently for that reason. Those are different budgets, different teams, and different timelines, and the category label covers both.

Forrester does not endorse any vendor named here, and tier placement should not be read as a recommendation to buy.