October 5, 2026

Positioning an AI startup when rivals use the same models

Position your AI product on the value it delivers against what buyers would otherwise do, for the customers who care most about that value. Don't position on the model. Your competitors can buy the same ones, and a better model from your provider is a better model for them too. Build the position with April Dunford's five components, pick a category frame that makes your strengths obvious, and base your differentiation on what a model upgrade won't hand a rival.

Then prove it with evals published alongside their method, case studies with a baseline and a clear security page. Test it in sales calls and win/loss interviews before you rewrite the homepage.

Why is positioning harder for AI products?

Most enterprise LLM spend goes to three providers. Menlo Ventures, an Anthropic investor, surveyed about 500 US enterprise decision-makers in November 2025 and put Anthropic's share of enterprise LLM spend at 40%, OpenAI's at 27% and Google's at 21%. The models are close, too. In Stanford's 2026 AI Index, the top models from Anthropic, xAI and Google scored 1,503, 1,495 and 1,494 on Arena, where people vote between two anonymous answers. Andreessen Horowitz investors wrote in 2023 that many generative AI apps were "relatively undifferentiated" because they relied on similar models and hadn't found network effects, data or workflows that competitors would struggle to copy.

Upgrades reach your rivals when they reach you. In Menlo's mid-2025 survey, 66% of builders had upgraded models within their existing provider over the past year, and only 11% had switched vendors. Dunford wrote in September 2026 about a founder who said Anthropic had launched three major releases that summer and that any positioning would be out of date in two weeks.

Features get matched in weeks. Google launched Deep Research in Gemini on December 11, 2024. OpenAI launched deep research in ChatGPT on February 2, 2025, and Perplexity launched Deep Research on February 14. If a platform can ship your headline feature within a quarter, it can't be your position.

The real alternative is often an AI tool your buyer already uses. In Microsoft and LinkedIn's 2024 Work Trend Index, 78% of knowledge workers using AI brought their own tools. And the word AI can work against you. In Washington State University experiments with more than 1,000 US adults, mentioning AI in a product description lowered emotional trust and purchase intent, more so for high-risk products. They were consumers, not software buyers, so treat it as a warning. Lead with the job and the result.

Start with April Dunford's five components

April Dunford's Obviously Awesome is the method I'd start with, and she released a second edition in February 2026. Her 2021 quickstart guide lists five components: competitive alternatives, differentiated features or capabilities, value for customers, target customer segmentation and market category. Each depends on the others, so start with the alternatives and choose the category last.

Competitive alternatives. List what customers would do if you didn't exist, which can be nothing. For AI products, that's usually a person or agency, a general assistant, an in-house build or the status quo. Menlo found 76% of enterprise AI use cases were bought rather than built in 2025, so about a quarter are still built in-house. Count only what buyers shortlist. "They are not a true alternative if they don't show up on a shortlist," Dunford wrote in 2024. In May 2026 she advised positioning against the competitors prospects shortlist today, not competitors who might show up later.
Distinct capabilities. List what you can do that each alternative can't. Beating an agency on turnaround is a different capability from beating a general assistant on accuracy.
Value. Turn each capability into what it gets the buyer, grouped into two or three themes. Dunford's advice to AI founders is that value themes often evolve more slowly than features.
Best-fit customers. Her 2021 guide calls them "customers that really care a lot about your unique value." Look for the trait that makes your value matter, such as task volume, the cost of an error or the data they hold.
Market category. The frame that makes your value obvious to those customers. More on that next.

Should you pick an existing category, a subcategory or a new one?

Your category sets expectations before buyers see a demo. Dunford's 2021 guide says positioning context "sets off a really powerful set of assumptions about who your product competes with, what features your product should have, who the product is intended for, and even things like what the product should cost." So your category and your pricing model have to agree.

In a 2019 post she describes the options: take on the leaders of an existing market head to head, split off a sub-segment you can own, change how people think about the category, or create a new one, which she calls the most difficult option.

Choice	When it fits	Trade-off
Existing category	You beat the leaders on what buyers already value	Instantly understood, but compared feature by feature with incumbents
Subcategory	A segment has needs the leaders serve badly	Borrows the category's demand and budget, but caps your early market
New category	No existing frame makes your value obvious	You own the name, but nobody searches for it yet

In 2017 she wrote that in a new market people aren't searching for solutions, so advertising and SEO have limited success. Her 2021 guide adds that category creators often lose in the long run to companies that arrive after the hard work is done. A new name can also give AI assistants nothing to match when a buyer asks for the best tool for a job. For most AI startups, I'd pick a subcategory and make the qualifier about the buyer or the job, since every rival can claim "AI-powered." As an illustration, "accounts payable automation for multi-entity construction firms" says who it's for and what it replaces.

What differentiation survives the next model release?

Test each capability against one scenario: every competitor gets the next model tomorrow. A prompt, a model choice or a feature a platform could ship fails that test. These hold up better.

Workflow depth: the approvals, exceptions, handoffs and audit trail around each model call.
Proprietary data from real use, if your contracts allow it. Buyers will expect the default OpenAI and Anthropic give business customers, which is no training on their data, so get explicit consent.
Integrations: two-way connections into the systems your buyer runs.
Distribution: partnerships, marketplaces, a community or a founder audience. Features copy faster than channels.
Trust and security: certifications, data controls and a record with careful buyers.
Evaluation results on your buyers' tasks, rerun on every model change.

When a new model does change what you can do, rerun the first three components. Dunford's September 2026 post argues that positioning built on a feature list will feel like it needs changing every time the product does. Change it when the value changes.

What proof does a skeptical buyer need?

In Menlo's 2024 survey of 600 US IT decision-makers, measurable value (30%) and understanding the context of their work (26%) far outranked the lowest price (1%) as selection criteria. Data privacy hurdles came up in 21% of failed pilots. You need proof for both.

Evals with the method attached

A benchmark number without a method is just a claim. The FTC's substantiation policy expects a reasonable basis before a claim runs. In April 2025 it alleged a company had promoted its AI content detector as "98 percent" accurate while independent testing showed 53% on general-purpose content. The final order, approved in August 2025, requires competent and reliable evidence for such claims. Publish these with every result:

The task set, its size and source, and whether you tuned on it.
The baseline, meaning the buyer's real alternative, such as an expert or a general assistant with a good prompt.
How outputs were graded, and how you checked the grader.
Your release, the underlying model and the date.
Where the product fails.

Sierra's research team published τ-bench in June 2024, with a paper, code and data for testing customer service agents. Four months later, Anthropic reported Claude's τ-bench scores in a model launch. Independent tests carry more weight. The Vals Legal AI Report (February 2025) tested four legal AI products on seven tasks against lawyers who didn't know they were in a study.

Case studies a buyer can check

Name the customer, task, baseline, period and metric, and say what else changed. If the customer won't be named, describe the segment and offer a reference call.

A security and data-handling page

Answer procurement's questions before they're asked.

Model providers and subprocessors, and where data is processed and stored.
Whether you or your providers train on customer data, with links to the terms you rely on.
Retention. OpenAI may keep API inputs and outputs for up to 30 days, and offers zero data retention on eligible endpoints for qualifying use cases.
SSO, roles, audit logs and deletion on request.
A SOC 2 report and, if buyers ask, ISO/IEC 42001, the AI management system standard published in December 2023.
How you handle wrong outputs, and who to contact.

How do you turn positioning into messaging?

Positioning is internal. Buyers see a messaging hierarchy built from it.

Point of view: what you believe your market will look like. Dunford argued in June 2026 that prospects want this now, rooted in what you do better than any other vendor.
Headline: category, best-fit customer and main value, in the buyer's words.
Two or three value themes, each tied to the alternative it beats.
Capabilities under each theme.
Proof next to each theme.
A next step that fits how your buyers buy, such as a sandbox, a demo or a paid pilot.

Check your homepage

Nielsen Norman Group wrote in 2011 that people often leave web pages within 10 to 20 seconds, and that a page must communicate its value proposition within 10 seconds to hold attention longer. Show your homepage for 10 seconds to five best-fit buyers, then have them tell you what it is, who it's for, what it replaces and why it's better. Next, delete "AI" from the hero. If the message collapses, you're selling technology. Finally, paste a competitor's name over yours. If every sentence is still true, you haven't positioned yet.

Check how AI assistants describe you

Ask ChatGPT, Gemini, Claude and Perplexity what you do, who you're best for and what your alternatives are, five times each because answers vary, then compare them with your positioning. If they put you in the wrong category or name alternatives you never meet in deals, fix your own pages first, then the third-party pages they cite. My guide to AI visibility for SaaS and AI startups covers the method.

How do you test positioning before rewriting the site?

Test where feedback is fastest. Dunford wrote in 2024 that she liked spending time on sales calls, first calls especially, to see how a message was landing.

Sales calls. Use the new positioning on your next 10 to 15 first calls. Note whether prospects repeat your category back and which value theme makes them lean in. If they keep asking what you are, the category is wrong.
Win/loss interviews. Talk to recent buyers within a few weeks, while they remember the shortlist. Ask what they used before, who else they considered and what tipped it. Dunford wrote in May 2026 that she finds wins more relevant to positioning than losses.
Landing page tests, once traffic allows. As an illustration, lifting a 2% demo rate to 3% at 5% significance and 80% power needs about 3,800 visitors per variation by the standard two-proportion formula. At 100 visitors a day split across both versions, that's about 11 weeks, so most early-stage sites should test in sales first.

A one-page positioning worksheet

Fill this in with your founders and your sales and product leads, and attach customer evidence to every answer.

Field	What to answer	Evidence
Competitive alternatives	What best-fit buyers would do without you, including nothing	Shortlists from your last 10 deals
Distinct capabilities	What you can do that each alternative can't	Demo, eval results
Value themes	What those capabilities get the buyer, in two or three themes	Customer quotes and metrics
Best-fit customers	Who cares most about that value, and what they share	Won deals, retention by segment
Market category	The frame that makes your value obvious to them	Words buyers use on calls
Point of view	Where your market is going, based on your strengths	Roadmap, founder writing
Model-upgrade test	Which capabilities hold if a rival gets a better model	Pass or fail per capability
Proof	What would convince a skeptical buyer of each theme	Published evals, case study, security page
Assistant check	How ChatGPT, Gemini, Claude and Perplexity describe you	Saved, dated answers

If you want help

Our product and GTM teardown ($2,000, two to three weeks) reviews your positioning, pricing, onboarding and go-to-market, and ends with a written plan and a readout call. See how we work with AI startups, or book a 30-minute call.

Sources

Facts checked on 4 October 2026.