// guide
Why is "which LLM do you use?" the wrong question?

Written by: Dmitry Filippov, Founder, GET-GEO.AI
Published: 2026-08-03 · Updated: 2026-11-18
// short answer
A model’s behavior depends on the system around it: search, tools, instructions, ranking and safety controls. The same model can therefore cite different sources in different products. GEO addresses how each product finds and reads pages, and which external sources support the information on them.
- “Which LLM?” is the wrong question: put one model inside two products, Cursor’s cloud agents and Claude Tag (Anthropic, June 2026), and the harness gives it two different jobs.
- A harness has four parts: the system prompt, the tools, context and memory, and the working loop; change them and the same model becomes a different product.
- Before paying for a stronger model, check the four harness parts: a model upgrade cannot add tools, memory or data access that the product does not give it.
- When an assistant searches, three steps run: a search tool finds pages, a parser extracts text, the model composes. In our September 2026 study, 22 of ChatGPT’s 52 answers carried no recordable source link.
- Measure each product of one vendor separately: across 20 Gemini app and Google AI Mode answers in our 14 September 2026 check, one complete source URL we recovered appeared in both. One day, incomplete URLs.
The same model can do different jobs
Businesses often start evaluating an AI tool by asking which model it uses: GPT or Claude, which version, and whether the latest release will make the product better. Those are reasonable questions, but they reveal little about the tools and information the product can use to complete a task.
Suppose the same model sits inside two products that both take requests in Slack. The first is Cursor’s cloud agents, a coding product: started with an @cursor mention, an agent runs in an isolated virtual machine on a model you select from Cursor’s list, clones the repository, writes and tests the change, and can open a pull request.
The second is Claude Tag, which Anthropic launched in beta on 23 June 2026: Claude joins selected Slack channels as a team member, builds context from what it reads there and connects to the tools, data and codebases an administrator chooses. It can write code too — but in a channel set up for sales work it could just as well review a client thread and suggest how to answer a third request for a discount.
In that scenario the underlying model is the same, but the products give it different roles, tools and context. A coding agent may be able to discuss a customer conversation, yet it will lack the information and workflow of a tool designed for that job.
What a harness is
A harness is the software and configuration around a model: the instructions, tools, context and workflow that turn it into a usable product.
- The system prompt — instructions that define the role, tone, priorities and boundaries. It decides whether an agent reduces every task to code or reasons about the business.
- Tools — what the agent can access. A terminal and git make it a programmer. A CRM and email make it a sales assistant. Web search and a crawler make it a researcher that reads your site and your competitors' sites before giving advice.
- Context and memory — what the agent knows about you. Channel history, company documents, past decisions. A model without context meets you for the first time in every conversation.
- The working loop — how the cycle is structured: a single reply, a long autonomous session, self-verification, escalation to a human.
Change the harness, change the product
Change the harness, and the same model becomes a different product: it gets another role, other tools and other information to work from. The reverse matters when a product disappoints. A model upgrade cannot add a tool, a memory or access to data that the product does not give it, so we advise checking the harness before paying for a stronger model. We have not benchmarked models against harnesses ourselves; treat this as a working rule and test it on your own task.
Why this concerns your website
All of this applies not only to the tools you buy, but to how AI sees your business. When a potential customer asks ChatGPT, Claude or Perplexity "which service should I choose for X," the harness first decides whether to search the web at all. In our study of the pages assistants cite, run on 14 and 15 September 2026, 22 of ChatGPT’s 52 answers (signed out, web search on) carried no source link we could record: either no search ran, or the interface showed source labels without URLs, and our figure does not separate the two. When a search does run, a search tool finds pages, a parser extracts text, and only then does the model compose an answer. At every step, your website either passes the filter or it does not.
Which product answers can also change where the sources come from. On 14 September 2026 we asked five GEO buyer questions twice each in the Gemini app and in Google’s AI Mode, two products of one vendor: across those 20 answers, one of the complete source URLs we could recover appeared in both (the check is described in our guide to getting cited by Gemini). On 27 September 2026 we put 30 local-business questions to ChatGPT twice each, across eight niches in Houston, Miami and London: in the 55 answers that showed a map, 280 of the 305 businesses it named were also on the map’s cards, a sign that a business-listings tool, a different part of the harness from web search, shapes that shortlist. Both are one-day checks in English. For three of the five Gemini questions complete URLs could not be recovered, AI Mode showed part of its sources as labels without URLs, and the check did not establish which model version answered in either product. Read it as a first look at two products listing different sources, not as one model behaving in two ways.
A beautiful site built on heavy client-side JavaScript may be invisible to the parser. A page without clear structure — headings, facts, specifics — can be read and still not cited: there is little for the model to lift. With few mentions on external sources, the search step has little besides your own pages to find. That can be enough for a narrow question: in six clean ChatGPT sessions on 13 and 17 August 2026, questions about a multilingual GEO agency put us first each time, and the 17 August answer cited our site inline. That is one engine and our own brand, on narrow questions, and the citations in those runs also included our Crunchbase, LinkedIn and Trustpilot profiles. For a broad "best X" question, plan for third-party pages to carry part of the answer: on the vendor-choice questions in the September study above, directories and comparison sites took 77 of the 306 links (25%) across its four assistants.
This is exactly what GEO (Generative Engine Optimization) is about: optimizing not for a ranking algorithm, but for the harness of generative systems — for how AI agents discover, read and retell your content. Classic SEO asked "how do we get into the top 10 links?" GEO asks "how do we get into the answer itself?"
What to do with this
When choosing an AI tool, check whether its tools, data access and workflow fit the task. These can matter more than whether it uses the most capable model.
When building an internal AI tool, first define the agent’s role, what information it can see, what actions it can take and where it must stop. Choose the model within that design; it can be changed as requirements develop.
For website visibility, consider what an assistant can retrieve and use from your pages. Clear, accessible information gives it useful context when answering a customer’s question. Check each assistant separately and ask each question more than once: as the checks above suggest, a result in one product may not carry over to another. We measure four engines per language: ChatGPT, Perplexity, Gemini and Google AI Overviews or AI Mode where offered.
Related questions
What is a harness in AI products?
The harness is everything wrapped around a language model that turns it into a working tool: the system prompt, the tools it can call, the context and memory it has about you, and the working loop that structures how it acts. Change the harness and the same model becomes a different product.
Does upgrading the model fix a weak product?
Not by itself, when the weakness is in the harness. A stronger model cannot call tools the product does not give it or read data it has no access to. We advise listing what the task needs (tools, memory, data access, a working loop), fixing what is missing, then comparing models on the same task.
How does the harness idea relate to GEO?
When ChatGPT, Claude or Perplexity answers a commercial question, its harness decides whether to search the web. When it does, search finds pages, a parser extracts text, and the model composes the answer. GEO prepares your site, and the outside sources that describe you, for those steps, so assistants can discover, read and cite you.
Should I ignore which LLM a tool uses?
No — the model still matters. Treat it as one input among several. Ask first whether the harness matches your task: role, tools, access to your data, and how the working loop is designed. The model can often be swapped later, so we advise settling the harness first.
Related guides
Sources
- 01Anthropic — Introducing Claude Tag (23 June 2026)
- 02Anthropic — Claude Managed Agents overview (how Anthropic uses the term “agent harness”)
- 03Cursor — Cloud Agents documentation (formerly Background Agents)
- 04Aggarwal et al., GEO: Generative Engine Optimization (2023), arXiv:2311.09735
- 05Our research: what kinds of pages AI assistants cite (569 links, four assistants, 14–15 September 2026)
- 06Our research: where ChatGPT gets business ratings (map and rating check, 27 September 2026)
- 07Our guide: how to get cited by Gemini and Google AI Mode (includes the Gemini app and AI Mode check of 14 September 2026)
- 08Our case study: six clean runs on multilingual GEO queries (13 and 17 August 2026)
// share
How to cite this page
You may quote and reuse this content with attribution and a link to this page.
“Why is "which LLM do you use?" the wrong question?” — GET-GEO.AI, 2026-11-18. https://get-geo.ai/en/guides/model-vs-harness