Product Role Navigator

The obvious thing would be to score all twelve roles from that framework's own role table instead of by hand. It does not work. It covers two of them, has no product owner, technical PM, growth PM or AI role at all, and states its levels as cumulative minimums per seniority tier — so a framework-derived Head of Product comes out a flat 5 on every axis. Profiles built that way nest inside one another instead of differing in shape, which is precisely the differentiation a radar exists to show.

The two sources also answer different questions. These scores describe what a role leans on day to day: a Head of Product is low on Process because they do not run ceremonies, not because they could not. The framework describes the floor someone must clear to hold the job. Both are useful; averaging them would be neither. Where the framework does speak to seniority, it appears in the Level guide as "Typical role at this level", read straight off the same framework table.

The 2026 taxonomy describes all twelve roles in words — what each focuses on, how far ahead it plans, who it works with and how it is judged — but assigns no numbers. Seven roles carry scores from the original brief. The other five (Product Marketing Manager, Platform Product Manager, AI/ML Product Manager, Data Product Manager, AI Product Designer) are badged derived: their six values were taken from that description, and each one shows the reasoning in its role detail so you can disagree with it. Nothing here is measured.

Every role can be re-scored by hand, from Edit scores at the bottom of Role detail. An edited role is marked updated in the legend and the role list, its badge changes from Reference to Edited by you, and its original values are shown underneath so the change stays legible. Edits are yours alone — they are stored in this browser and travel with neither the file nor a link. Revert to original restores a single role; Reset all appears in the role list once anything has been changed.

Leaving a field empty marks that axis as not applicable to the role. It is then excluded from the fit calculation rather than counted as zero, and the role is drawn as an open shape rather than closed through a value it does not have.

Five bands, because two independent frameworks that describe professional competence both use five: the Dreyfus model of skill acquisition, and the European e-Competence Framework (EN 16234-1), whose levels e-1 to e-5 map approximately onto EQF levels 3–8. The e-CF grades its levels on three dimensions: autonomy (from responding to instructions to making personal choices), context complexity (from structured and predictable to unpredictable and unstructured) and behaviour (from able to apply to able to conceive).

That framework has four ascending levels, not five. Bands 2–5 line up with it and band 1 sits below its lowest descriptor. That offset is this tool's assertion — neither framework states it. Everything inside a band is quoted; only the alignment is ours.

• A1 Strategy & Vision → Strategic ownership
• A2 Execution & Delivery → Product management
• A3 Data, Market & User → Applying user-centred insights
• A4 Communication & Alignment → Stakeholder relationship management
• A5 Process & Agility → Agile and Lean practices
• A6 Technology → Systems design + Technical design throughout the life cycle
• A7 Design & UX → Iterative design + Evidence-based design
• A8 Working with AI systems → nothing. No published framework describes product work with non-deterministic systems; the nearest are data-scientist skills about building models. Its level descriptors are written for this tool and say so wherever they appear.

The "Typical role at this level" line is read directly off that framework's product manager role table: it names the role levels that require exactly that skill level. A1–A5 map to product manager role levels; A6, A7 and A8 do not, because the framework does not require those skills of product managers.

The axis set changed. Technology and Design were one axis until they were split: the technical half rested on the framework's generic four-bullet descriptor welded to a designer's skill, and a posting strong on one and silent on the other produced a meaningless middle score. Splitting them made both properly sourced — 26 and 27 descriptor bullets against the old four. The AI axis was added at the same time, and is the only one on the page with no published source behind it. The re-scoring of all twelve roles across those three axes is ours, not a source's.

Years of experience is the weakest signal on this page, and the one thing here no framework will stand behind. SFIA states it plainly: "Time alone does not determine competency… Sustained performance over time is evidence of reliability, not a proxy for competence in itself." The Dreyfus model has no empirically validated timeframes for stage progression. Published product career ladders disagree with each other — one widely cited set puts Associate PM at 0–4 years, Senior PM at 3–5, Principal at 4–8 and Director at 10+, ranges that overlap and contradict.

One axis does not use those bands at all. Nobody can have eight years of experience shipping products on language models, because the practice is only a few years old — so A8's bands are scaled to the age of the field and computed from ChatGPT's release rather than written down, which stops them going stale as the field ages. Related work existed before that, but classical machine-learning product work is a different craft from the one these descriptors describe.

The tenure bands shown elsewhere in the guide are an indicative aggregate of those published ladders, marked with a dashed rule to signal that they are softer than everything beside them. Use the task descriptors and the role anchor to decide a level. Use years to sanity-check it, never to set it.

Every number on this page is a level on the same scale, whether it describes your profile, a job, or a reference role. A profile at 4 describes someone who can work at that level; a position at 4 describes a job that needs someone who can. That is what makes subtracting one from the other mean anything.

This was not always true. The advert analyser originally scored how prominent an axis was in the text — "a major focus of the job" — while profiles and the level descriptors scored capability. The two were being subtracted from each other anyway. The analyser now scores the level a job needs, and is told outright that how much an advert talks about something is not the level it demands: one sentence asking you to own a three-year roadmap outranks three paragraphs about collaborating with designers.

The page works two ways from the same scores. Sizing up a role, you analyse an advert and place yourself against it. Hiring for one, you rate what the job needs and name which roles those requirements come closest to — a shortlist, because titles are not a partition and two roles are often within half a point of each other.

For a profile C and a role S, over the n axes both actually define: cumulative gap is Σ|Cᵢ − Sᵢ|, distance is √Σ(Cᵢ − Sᵢ)², and fit is max(0, 100 × (1 − distance ⁄ √(n × 4²))) — a largest possible distance of 9.80 across six axes, 8.94 across five.

Axes an advert never mentions are excluded rather than counted as zero, and the divisor shrinks with them. That has a consequence worth knowing: the fewer axes a posting speaks to, the higher any percentage looks, because the same distance is measured against a shorter ruler. A verdict is therefore withheld below five readable axes and below 60% fit.

It is also withheld for reasons that have nothing to do with the arithmetic. The analyser is asked three questions before it scores anything: whether the text is a job advert at all, whether the job is a product role, and whether the product is digital. A recipe, a careers article about a job family, and an advert for a heating engineer each fail the first two, and are shown and then dropped rather than loaded onto the radar. The third is subtler: a product manager for frozen food or for an insurance tariff is a real product manager, but these axes come from a digital-and-data framework and all twelve reference roles are software roles. Strategy, delivery, market and alignment carry across; technology, design and AI do not mean the same thing. Those readings are shown in full, without a percentage.

Each row also reports the cumulative gap and the widest single axis, because the percentage alone cannot separate "off by a little everywhere" from "off by a lot in one place" — both score 75%, at gaps of 6.0 and 2.45. RMSE is deliberately absent: it works out to 4 × (1 − fit/100), so it only restates the percentage.

1. Government Digital and Data Profession Capability Framework — Product manager, UK Government. Role levels and per-skill level requirements. Last updated 28 February 2025.
2. Government Digital and Data Profession Capability Framework — skill descriptions export (CSV, July 2026). Source of every "you can" bullet, quoted verbatim. Crown copyright, Open Government Licence v3.0.
3. Dreyfus, S. E. & Dreyfus, H. L. (1980), A Five-Stage Model of the Mental Activities Involved in Directed Skill Acquisition, University of California Berkeley Operations Research Center, for the USAF Office of Scientific Research. Via Wikipedia's summary.
4. European e-Competence Framework (e-CF), EN 16234-1 — five proficiency levels graded on autonomy, context complexity and behaviour; approximate mapping to EQF levels 3–8.
5. SFIA — Knowledge, skill and competency. Source of the statement that time alone does not determine competency.
6. Ravi Mehta, The Product Competency Model (Product Execution / Customer Insight / Product Strategy / Influencing People). Corroborates the axis taxonomy; defines no proficiency levels of its own.
7. Taxonomy of Product Roles in Software Development — A Comprehensive Competency and Operational Matrix (2026 Edition), supplied with this project and kept at docs/. Source of all twelve role descriptions: focus and horizon, operational scope, key interfaces and success metrics. It assigns no numeric scores.

Everything stays in this browser, in localStorage, under names starting prn: — both profiles, the current mode, which panels are open, your edited scores, which role groups are expanded, the theme, whether you have seen the first screen, and your Gemini key and model if you added them. Nine keys at most. No cookies.

Measured, not estimated. On a fresh browser that has just rated a position on all eight axes in whole numbers, three keys exist: both profiles, the mode, and the first-screen marker. They hold 231 bytes of values, or 262 bytes counting the key names, and 221 of that is prn:profiles. That is the floor. Half-point scores, naming a position, opening panels, editing role scores and adding a key each add more.

Two things leave only when you ask. Your Gemini key is stored as plain text in prn:aikey; this page never encrypts it, so treat this browser the way you would treat an unlocked password manager. And a Share link puts both profiles in the web address itself — eight scores each, plus anything you typed: the name on your profile, the position's title, and the company if an advert supplied one. Anyone with that link has all of it.

One thing leaves whether you ask or not: loading this page. The server that hosts it sees your IP address and which page you requested, as every web server does. That is the price of the tool being a link. Run the downloaded copy from your own disk and it does not happen — the tool works the same either way.

And one more, only if you have not added a key of your own. The AI features then run on a shared key, and the text you paste — your CV, or an advert — goes to this site's server, which passes it to Google. It is not stored: the request is forwarded and forgotten. Those calls use the owner's Google account, so the owner's quota and terms apply rather than yours, and there are three requests every ten minutes for everyone to share. Add your own key and none of this applies: the text goes straight from this browser to Google, and your key never reaches us.

The AI calls are: reading a CV, analysing an advert, testing a key and listing models. Each sends one request to generativelanguage.googleapis.com with your key in a header — never when the page loads, only when you press the button. The prompt is not a secret either: Copy the prompt in the analyser shows the exact text before you send it.