Model routing: definition and why AI answers vary

99Visibility Research · Updated October 21, 2026 · 4 min read

Short answer

Model routing is when an AI app sends each request to one of several underlying models instead of always using the same one. The app, the user's plan, or a menu choice decides which model answers. Routed models can have different knowledge cutoffs and different search behavior. So one app is not one knowledge state, and two people asking the same question about your brand can get answers from different models.

Key takeaways

Model routing means one AI app can answer with several different models, chosen automatically, by plan, or by the user.
Routed models can carry knowledge cutoffs months apart, so the same app can know different things about your brand.
Whether the answer used live web search is a second variable on top of the model, and it also changes per request.
Any test of an AI app should record which model and which search state produced the answer, or the result is an unknown mixture.
Nobody outside the vendor controls routing, so the useful moves are making your facts reachable and stated plainly on pages crawlers can fetch.

Model routing, precisely

Model routing is the step where an AI product decides which model handles a request. The product name stays the same. The model behind it does not.

Routing shows up in three forms:

Automatic routing (an auto router). The app picks a model for you. Secondary reports from September 2026 describe Perplexity's free tier as "Best", an auto-router.
Plan-based routing. Your subscription decides the model. Secondary reports from the same period map ChatGPT Free and Go to one model and Plus to another. OpenAI has not confirmed that mapping itself.
User model selection. You pick from a menu. Perplexity's paid tiers and Claude.ai both offer one, per the same September 2026 snapshot.

Most real products combine more than one of these.

How routing changes the answer

Each model has its own training data, frozen at its own knowledge cutoff. OpenAI's model documentation, checked in August 2026, lists a February 16, 2026 cutoff for the GPT-5.6 family. The Instant path behind ChatGPT's chat-latest alias lists August 31, 2025. That is about five and a half months apart inside one app.

Search behavior sits on top of the model. As of October 2026:

ChatGPT decides per query whether to search, and OpenAI says it rewrites your prompt into one or more targeted queries when it does.
Claude.ai web search is a user toggle with no published default.
On Perplexity's Agent API, web search is opt-in.
Google says AI Overviews and AI Mode "may use different models and techniques" and can return different links.

So every answer is a combination of a routed model and a retrieval state. The differences between free and paid ChatGPT are one visible result.

Why it matters for AI visibility

When search is off, the model answers from memory. Nothing you published after its cutoff can appear. When search is on, your page has to be fetched and retrieved before it can be cited, which is a separate pipeline covered in how AI search picks sources.

This means a single "what ChatGPT says about us" check describes one model in one state. Say you run a dental clinic in Austin and changed your prices in March 2026. A model with an August 2025 cutoff, answering without search, can still quote the old prices.

What to do:

Record the plan, model and whether citations appeared every time you test an answer.
Compare answers across plans before concluding anything. Switching models can change what the AI says about your brand.
Make sure your current facts sit in server-rendered HTML that retrieval crawlers are allowed to reach, so search-on answers can pick them up.

The free checker on this site gives a snapshot for up to three buyer questions. Treat it as one sample, not a fixed truth about every routed model.

Common misconceptions

"ChatGPT knows X about us." ChatGPT is a product, not one knowledge state. Different routes can hold different facts.
"An API test equals the app." A tool calling one API model with search off measures that model only. See whether tools test the same models people use.
"You can optimize for the router." No vendor publishes routing rules, and no published evidence shows a publisher lever on them. Access and clearly stated facts are what you control.
"The newest model always answers." Plan and routing decide that. Many users get an older or lighter path.

Frequently asked questions

Is model routing the same as model selection?

Model selection is one kind of routing, where the user picks the model from a menu. Routing is the broader term. It also covers automatic routers that pick for you and plan-based routing where your subscription decides. All three produce the same effect: the product name stays constant while the model and its cutoff change underneath.

Can I tell which model answered in ChatGPT?

Not reliably from a consumer answer alone. Plan-to-model mappings come mostly from secondary reports, and OpenAI has not published them in full. You can note your plan, any model the interface shows, and whether citations appeared. Visible citations suggest the answer used search. No citations means you cannot be sure which state produced it.

Does routing explain why two coworkers get different answers about my company?

Often, yes, along with other factors. Different plans can route to different models, and search fires at the model's discretion per query. Memory and custom instructions can also shape how ChatGPT rewrites a search, per OpenAI. Compare answers logged out with memory off before treating a difference as meaningful.

Can I influence which model an app routes to?

No. Routing is decided by the vendor, the user's plan, or the user. What you control is upstream: whether retrieval crawlers like OAI-SearchBot and PerplexityBot can reach your pages, whether your key facts appear in HTML before JavaScript runs, and whether those facts are stated consistently across your site.

Sources

This page is AI-assisted. It was written from a verified evidence base and checked against our editorial policy. Engine behaviour changes often; facts are as of October 21, 2026.