Most marketers need to understand 2 types of LLMs. The first is the general-purpose chat model, which answers everyday questions. The second is the reasoning model, which plans multi-step research before it answers. Between them, these 2 decide which pages AI assistants read and cite. Small and open-weight models matter if you build tools or work in a regulated field. Domain-specific models matter in those situations too. The big chat models now include multimodal ability, so treat it as a feature you write for.

Nobody keeps an official list, but marketers run into 6 types of large language models. 3 of them describe what a model does. Chat and reasoning belong to this group. Multimodal is the third. 2 describe size or training: small models and domain-specific models. The last one, open-weight versus closed, is about who can download and run the model. Because these labels describe different things, they overlap. One model can be a closed, multimodal chat model with a reasoning mode switched on. Read the categories as traits a model can combine, and expect a single product to have more than one.
| LLM type | What it is | Relative cost to run | Typical use | Strength | Limit |
|---|---|---|---|---|---|
| General-purpose chat | Broad model tuned to follow instructions | Medium | Everyday questions, drafting, search answers | Fast, broad knowledge | Training cutoff, can state errors with confidence |
| Reasoning | Model that works through steps before it answers | High | Research, comparisons, planning | Accuracy on hard tasks | Slower, uses more tokens |
| Multimodal | Model that takes images, audio or video as input | Medium to high | Image questions, document reading | Reads charts and screenshots | Mistakes when visuals are unclear |
| Small language model | Model with a few billion parameters | Low | On-device features, narrow tasks | Cheap and quick | Less world knowledge |
| Open-weight | Model whose trained weights you can download | You pay for your own hosting | Custom tools, private data | Control, fine-tuning | You run and secure it |
| Domain-specific | Model trained or tuned on one field's data | High to build | Finance, law, medicine | Field accuracy | Narrow, costly to maintain |
General-purpose chat
Reasoning
Multimodal
Small language model
Open-weight
Domain-specific
When you read a vendor's spec sheet, check which traits it claims. "Multimodal" and "reasoning" are modes inside one product. Small and open-weight mean a separate model. Domain-specific does too.
A general-purpose chat model is a large model trained on broad text, then tuned to follow instructions and hold a conversation. It's the default engine behind most AI answers. It answers quickly because it writes its reply in one pass, word by word, without a planning stage.
Its knowledge stops at its training cutoff. So assistants add live web search when a question needs fresh facts, and that gives your brand 2 routes into an answer. The first is memory: what the model learned about you during training. The second is retrieval: the pages it fetches and cites at query time. Retrieval is the route you can change this quarter. Memory only changes when the model is retrained.
To test both, ask an assistant about your category with web search turned off, then ask again with it on. The difference between the 2 answers shows you how much depends on training data and how much on your current pages.
A reasoning model works through a problem in hidden steps before it writes the answer you see. It splits the question into parts and checks partial results. When a step fails, it goes back. The extra thinking costs money. OpenAI's API documentation says reasoning tokens are hidden from the user but billed as output tokens. PremAI, which publishes guides on AI infrastructure, estimates that thinking tokens can outnumber visible output by 5 to 20 times on complex problems.
For marketers, the bigger change is in search behaviour. A reasoning model doesn't run 1 search. It runs a batch of searches, one for each sub-question it has planned. Seer Interactive, a search marketing agency, measured 10.7 fan-out queries per prompt across 501 prompts on the Gemini 3 API, up from 6.01 with Gemini 2.5.
So check whether your page answers the sub-questions as well as the head term. A page about payroll software that skips pricing or setup time gives the model nothing to cite when it runs those searches.
A multimodal model takes in images alongside text and can also accept audio or video. It reasons across these inputs. When Google launched Gemini in 2023, it said it designed Gemini to be "natively multimodal, pre-trained from the start on different modalities." That means the model can read a chart or a product photo directly, with no separate text-extraction step in between.
How much this matters depends on your content. A software company's screenshots and a retailer's product photos now count as content the model can read and describe. Text baked into an image can be read too, but blurry or low-contrast graphics are where models misread figures.
Check that every chart has a caption and nearby text that states its main number. Keep alt text accurate. And don't put a key fact only inside an image.
A small language model (SLM) is a model with a few billion parameters, compact enough to run on a laptop or phone. The difference in the slm vs llm question is scale. Microsoft's Phi-3-mini has 3.8 billion parameters, and its technical report says it is "small enough to be deployed on a phone." The largest LLMs are orders of magnitude bigger.
Businesses are moving toward small models for narrow tasks. Gartner predicts that by 2027, organizations will use small task-specific models 3 times more than general-purpose LLMs. Sumit Agarwal, VP Analyst at Gartner, said these models "provide quicker responses and use less computational power, reducing operational and maintenance costs."
Small language models rarely write the AI search answers your customers read. Where you'll meet them is inside internal tools, such as tagging content at scale or sorting support tickets. Before you trust one with facts about the world, test it on your own data, because a small model holds far less general knowledge.
Open-weight models publish their trained weights, so anyone can download one and run it on their own servers. Closed models are reachable only through the vendor's app or API. OpenAI's help centre defines its gpt-oss release this way: the weights are available under Apache 2.0, which means "you can download the models, run them on your own infrastructure... and customize or fine-tune them."
Open-weight doesn't mean open source. You get the weights without the training data or code, and licences range from permissive to restrictive. The trade is control against effort. With a closed model, the vendor handles hosting and updates. With an open-weight model, your data stays on your own servers, but you pay for compute and handle security yourself.
Search marketers should care because companies build their own assistants on open-weight bases. If you're choosing one for an internal tool, read the licence for limits on commercial use before you build.
A domain-specific LLM is trained or tuned on one field's documents so it handles that field's language and tasks better than a general model. The best-known example is BloombergGPT, a 50-billion parameter model built from a 363 billion token dataset of financial documents plus public data. Gideon Mann, head of Bloomberg's ML Product and Research team, said: "The quality of machine learning and NLP models comes down to the data you put into them."
Few companies train a model from scratch now. Most fine-tune an existing model or connect a general model to their own documents through retrieval, which costs far less. Domain models show up in finance and law, where vocabulary is precise and an error has a cost. The same applies to healthcare.
If you market in one of those fields, your audience already asks questions in tools built on specialist models. Use the standard terms of your field consistently, because these models learned that vocabulary from industry sources.
Each type has well-known examples, and the leading business models come from a small group of vendors. Menlo Ventures' 2025 enterprise report estimates that Anthropic holds 40% of enterprise LLM spend. OpenAI follows at 27%, and Google holds 21%. For business use, then, Anthropic's Claude and OpenAI's GPT family are top models. Google's Gemini is another.
Model names change every few months. Check the date on this comparison and any other you rely on.
Each AI assistant runs on a model from one of the types above, a closed chat model with a reasoning mode attached. What matters for your visibility is which model each assistant uses and where it gets its web results.
ChatGPT uses a general-purpose chat model and a reasoning model, with a router that picks between them for each prompt. OpenAI describes GPT-5 as a system with a smart, efficient model for most questions and a deeper reasoning model for harder ones. The router chooses "based on conversation type, complexity, tool needs, and your explicit intent." It's closed and multimodal.
Google runs a custom Gemini model inside Search. In May 2025 it brought a custom Gemini 2.5 to both AI Mode and AI Overviews in the US. Google says AI Mode uses query fan-out, "breaking down your question into subtopics and issuing a multitude of queries simultaneously."
Perplexity's in-house Sonar model is a fine-tuned open-weight model. The Decoder reported that Perplexity built Sonar on Llama 3.3 70B and trained it further for factual accuracy and readability.
Claude is closed, and its web results come from a different index than Google's. In March 2025, TechCrunch reported that Anthropic added Brave Search to the list of partners that process Claude data. Microsoft Copilot's Smart mode is powered by GPT-5, according to Microsoft Support.
Model types differ in how they pick sources, but a page gets cited when it answers the specific sub-question the model searched for. A high ranking on the head term isn't enough. Ahrefs found that only 38% of AI Overview citations came from pages in the organic top 10, down from about 76% in its July 2025 study. The rest were spread across lower positions and pages outside the top 100.
For general chat models with search, write a direct answer in the first sentence of each section, because models quote passages. For reasoning models, cover the related questions a buyer asks next, like cost or setup time, so the sub-searches land on your page. For multimodal models, put the numbers from your charts into the text as well.
Assistants that don't use Google's index need separate checks. If Claude pulls from Brave, a page Google ranks well can still be missing from Claude's results. I'd test the same 10 buyer questions in each assistant every month and note who gets cited. The pattern shows you which model type you're losing to.
If you work in marketing, focus on general chat models and reasoning models first, because they write the answers your buyers read. The other types of LLMs matter when you build internal tools or sell into a specialist field. To measure where you stand, run your top buyer questions through ChatGPT and Google AI Mode. Test them in Perplexity too, then record which pages each one cites. Repeat after every major model release, since a new model can change its sources overnight. Snoika audits websites and tracks brand visibility in Google and AI assistants. Check your AI visibility free.
There isn't an official count, but this guide covers six overlapping types of llms. The labels describe different traits, so a single model can belong to several categories. Count them as ways to compare models rather than separate alternatives.
The best fit depends on the task: general-purpose chat models suit everyday work, while reasoning models suit complex research. For narrow internal tasks, small models can reduce running costs. Compare candidates on your own data rather than choosing by popularity.
Open-weight models let you download and run the trained model yourself, while closed models are accessed through a vendor's app or API. Self-hosting means taking responsibility for compute and security. Check the licence before commercial use, because open-weight doesn't necessarily mean open source.
A chat assistant can combine a general-purpose chat model with a reasoning model, selecting between them according to the prompt. It can also support images. The chat interface alone doesn't tell you which model or mode handles a particular answer.
Compare billed usage and hosting responsibilities, rather than the advertised model price alone. Reasoning models can charge for hidden reasoning tokens as output. If you host an open-weight model yourself, your budget also needs to cover compute and security.
Keep reading
All articles


See how AI talks about your brand