Methodology

BakeHQ is a set of typed entities connected by an explicit graph, not a collection of articles. Products, ingredients, ingredient functions, techniques, science concepts, faults, causes, corrective actions, equipment, comparisons, guides, glossary terms, history topics and tools are each modelled separately, with their own fields.

Some of those types have no pages. A cause, a corrective action and an ingredient function are real nodes carrying real content, but none is something a reader would search for on its own — so they render inside the entity that gives them meaning rather than at a URL of their own. That is a deliberate position, not an unfinished one.

Every connection between two records is a hand-authored edge with a stated predicate, its own confidence, its own sources and a mandatory note explaining why the connection exists. That note is rendered beside the link.

Nothing is inferred from name or tag similarity. An unexplained “related to” link is the most common way a knowledge site pads its internal linking without adding information, and if there is nothing to say about why two records are connected, the connection is not authored.

Claims are placed on a six-rung ladder. The tiers exist because baking claims fail in distinguishable ways, and collapsing them is how a reference site ends up asserting kitchen folklore in the same voice as thermodynamics.

Established science

Measured and reproducible food science, described in textbooks or peer-reviewed work. The mechanism and the numbers are both supported.

Professional consensus

Agreed by professional bakers and baking-science texts, with an accepted mechanism, though not isolated by controlled experiment.

Common practice

Widely done and usually works. The reason it works may be plausible rather than demonstrated, and the effect size is often unmeasured.

Contested practice

Experienced bakers genuinely disagree, and there is reasonable ground on more than one side. Recorded as a disagreement rather than resolved by assertion.

Traditional account

A long-standing account of how something originated or was done. Recorded as tradition, which is not the same as recorded as fact.

Commonly repeated, but not supported

Frequently repeated and not supported by the evidence consulted. Recorded so that the correction is findable by someone searching for the claim.

There is no numeric confidence score anywhere in BakeHQ and none may be added. A single number invites false precision and conceals exactly the uncertainty this model exists to surface.

A fault records the symptom, the possible causes, and — crucially — the discriminating observations that tell those causes apart. Each answer names the causes it supports and the causes it rules out.

Elimination is treated as hard and support as soft: an eliminated cause is gone, while support only orders the survivors. The engine deliberately publishes no probability, because it has no basis for one — it does not know the base rate of each cause in your kitchen. It can also conclude that no cause matches, which is the honest outcome when the model has a gap.

Expansion is meant to be driven by real search demand rather than by a list somebody thought of. In August 2026 BakeHQ ran its first demand study: 290 question-shaped prefixes against Google’s public autocomplete endpoint, producing 2,635 distinct observed suggestions, which were reduced by published rules to 2,137 canonical intents across 28 opportunity clusters.

None of that is search volume, and BakeHQ holds none. Autocomplete shows that a completion is offered and roughly where Google ranks it against its siblings. Turning a rank into a monthly figure would be fabrication with an arithmetic step in front of it, so the evidence model gives ranked observation its own category with no field a number could be written into. Only two of the six evidence categories can carry a figure at all, and both require a named provider, a metric, a geography and a retrieval date.

Provider volume.

May carry a number, and only with full attribution. Nothing in the register currently uses this.

Relative trend index.

May carry a number, and only with full attribution. Nothing in the register currently uses this.

Autocomplete-observed.

Carries an observed rank and the prefix that produced it. This is the basis of almost the whole register.

SERP-observed.

Seen in a real results page.

Manually identified.

An editor's judgement. Honest, and weak.

Hypothesis.

A guess worth testing, and labelled as one.

Clusters are scored on ten factors, each rated 0–3, each weighted. The weights are published here rather than kept internal, because a prioritisation model nobody can inspect is an opinion wearing arithmetic.

Nine factors are editorial judgements. The tenth — demand strength — is derived from the evidence rather than authored, because it is the one an optimistic editor would inflate to justify work they already wanted to do.

Weights total 21, so the maximum score is 63. Competitive weakness is weighted low on purpose: a weak competitor is an opportunity, not a reason to build.

The register itself is not published. A public page of search phrases is keyword stuffing with extra steps, and BakeHQ would rather show the method than the keyword list.

Schemas validate each record in isolation at build time. These further rules validate what a schema cannot see, and the build fails on any violation:

BAKE-V001.

Every source reference resolves to a source in the register.

BAKE-V002.

A claim marked known with no sources must explain itself in a note.

BAKE-V003.

The established_science tier requires a qualifying source type.

BAKE-V004.

A source record carries citation metadata, never source material.

BAKE-V005.

A republished, aggregated or translated source names its original.

BAKE-V006.

Both sides of a comparison resolve to real, comparable entities.

BAKE-V007.

A comparison contains at least one material dimension.

BAKE-V008.

Every relationship predicate is used by at least one seed record.

BAKE-V009.

Every observation discriminates, and names only causes linked to its fault.

BAKE-V010.

Every cause of a fault has at least one corrective action.

BAKE-V011.

A search intent's declared target entity exists.

BAKE-V012.

A demand figure names a real provider, metric and retrieval date.

BAKE-V013.

No product slug collides with a reserved route segment.

BAKE-V014.

A published tool states its assumptions and its limitations.

BAKE-V015.

Every search intent's demand cluster exists.

BAKE-V016.

Every demand cluster has at least one search intent behind it.

BAKE-V017.

A cluster proposed for the next sprint implies actual work.

BAKE-V018.

Competitive posture and competitive-weakness score agree.

BAKE-V019.

An unsourced claim declares why it carries no source.

BAKE-V020.

A structural requirement is not qualified by a gas-producing mechanism.

BAKE-V021.

Sufficiency qualifies a performance, and hydration edges declare it.

BAKE-V022.

A mechanism corrected against evidence is not reasserted.

BAKE-V023.

An a-vs-b comparison resolves to the entities its slug names.

BAKE-V024.

A fault symptom is a folded, product-free description of an observed state, with honest provenance.

BAKE-V025.

A historical record states no precise date or inventor without marking the account as unsettled.

Baking is a domain where one word means two things across the Atlantic and both readers are certain they are right. BakeHQ models aliases with the register they belong to rather than picking one term as canonical, because a reader searching for either means something real and neither is using the word incorrectly.

Where two terms are not interchangeable — a British biscuit and an American biscuit are different products — the alias carries a note saying so.

See sources and evidence for the citation register and what each verification status honestly means, and editorial policy for how content is written and reviewed.