Methodology

How entries are collected, classified and dated.

Sources

Daily and weekly source monitoring. Archives are read for history. The declared schedule is separate from how often the vendor publishes. See Sources for each source's schedule, last attempt, last successful read, detected changes and dated health assessment. Coverage measures the health of vendor sources monitored daily or weekly. Historical archives and the announcements lane are excluded. Announcements made only in press releases are taken from the owner's weekly MarTech Watch and are linked to the vendor's own page. Every entry links to public evidence: product changes always use vendor sources (release notes, documentation, the vendor's newsroom or press releases).

Each cited vendor page is also saved in the Internet Archive's Wayback Machine, an independent copy the record does not control. Where one exists, the record links it as "archived copy" with the date the copy was taken: the earliest copy taken from the day before the state was first recorded onwards. Pages that the Wayback Machine cannot capture, such as some pages rendered in the browser or behind bot checks, have no archived copy. Cited links are checked weekly; a page that has been removed, or that redirects elsewhere, on two checks in a row is marked "no longer available at the vendor". The record itself does not change when a source disappears, and a check confirms only that the page is reachable, not that its content is unchanged. The project keeps its own copy of the text it read, which is not published.

Signals normally use vendor sources. One approved exception, the Hightouch offer for LiveRamp assets, links to Axios reporting and the co-CEO's public confirmation on LinkedIn because no vendor page is available. This exception applies only to that signal.

Automation

A language model reads the changed parts of each source and proposes entries in a fixed format. Every proposal is checked by an automatic validator (schema, taxonomy, date and state rules, allowed source domains) before it is merged. Entries the model is unsure about are published but marked under review and left out of the Release Matrix and side-by-side views.

AI-generated text label

Pages that show entries carry the EU icon for AI-generated content with a short label, and so does the footer of every page (EU AI Act, Article 50). What the AI does, what it does not do and how its output is checked: How AI is used.

What is checked, and what is not

Automatic checks on every record: data format; a capability from the current taxonomy; dates consistent with their precision (a month date is the 1st); states only move forward, never twice on the same day; promises and end dates consistent with their state; source on an allowed vendor domain (a press-wire copy raises a check that the vendor is the issuer); the date the record was first acquired matches the first fetch of its evidence. A record that fails is not published.
Under review means someone or something raised a doubt: the extraction (for example, no availability label in the source), a pipeline rule (an impossible state order, an announcement with no stated date, two sources that disagree on a date) or the owner. It is not a failed verification: the record keeps its source and stays out of the Release Matrix and side-by-side views until the doubt is resolved.
Where each state was read is shown next to its source: vendor release notes, changelog, release feed, community release post, blog, documentation, press release, or an announcement relayed by the author's MarTech Watch with a link to the vendor. This describes the source, not how reliable the entry is.
No per-record human verification. The owner reviews in batches and every resulting change is listed in Corrections; a record without corrections has not necessarily been checked by a person. The extraction model's own confidence score is not published: it is a self-assessment, not a calibrated measure.
Accuracy audit (9 October 2026). 50 records drawn at random from the 1,852 not under review, spread across all 15 platforms, were checked one by one against their cited source: all 50 exist there with the recorded lifecycle state; 50 of 50 dates are right (the audit first reported one exact day as unsupported; on 12 October that finding was found to be wrong, since the source gives the day, and the change was reversed); 49 of 50 summaries are fully supported by the source text (one adds a detail the captured page does not show). The audit will be repeated on a fresh sample.
Recall is a separate question: when the record started, a check against the author's MarTech Watch over 20 weeks found 78% of the Watch's items (88% including partial matches); the gaps were mostly press-only announcements and minor items, since covered by the announcements lane and the minor-update rule.

Lifecycle states

Announced: made public by the vendor, but the cited source does not show that it can be used yet (including press announcements that state no availability).
Preview: limited or early access, private beta. Beta: public beta.
GA: generally available; GA (regional) when limited to some regions or data centres.
Deprecation announced and Retired.

States only move forward. Preview, beta and GA come from the vendor's own product sources: release notes, changelogs, documentation, product feeds and community posts, or a dated vendor post that states availability. Press releases and announcements relayed by the MarTech Watch record Announced, unless the vendor's own page states availability (for example "now generally available"), in which case the state follows it.

Dates

The date is the vendor's own. Its precision is shown as published: a day, a week, a month (for example monthly release notes) or a quarter. A month-precise change is filed on the 1st of its month, so time windows such as "last 7 days" catch it only when that 1st falls inside the window. Separately, the pipeline records when it first saw each change, which measures its own delay; changes found in history published before a source was first read are marked as recorded from the source history.

“Confirmed by” means the state was reached on or before the evidence date; its start is not dated. A vendor promise is assessed against its named target state and the end of the promised period. Promises without a target remain visible without a score. Confirmation dates do not establish lateness. A deprecation end date does not establish removal: only a separately recorded retired state does.

Time windows on the site (last 7, 14 or 30 days) end on the data date, the UTC day of the latest dataset, and include both ends: "last 7 days" means that day and the six before it. Every page uses the same rule, and each window shows its dates. In the timeline, day-precise changes are grouped into dated weeks; week-, month- and quarter-precise changes have separate headers with the vendor's own precision. A dated week belongs to the archive month of its Monday; coarser dates belong to the month where their published period starts. No week is inferred from a month- or quarter-precise date. “Last 12 months” counts calendar months: the current month, even if partial, and the eleven before it.

Announced → available

The gap from announcement is measured to two milestones: first access (preview, beta or GA, whichever comes first) and GA. A gap in days is exact only when both dates are precise to the day and neither is a “confirmed by” date. Other pairs show a range or an upper bound and are left out of medians. A median requires at least five exact pairs, with no platform contributing more than half of them; otherwise the individual pairs are listed. Pairs whose dates are out of order are listed for checking, never counted as zero. Entries under review are left out.

Classification

Features are classified by what they do, using a capability taxonomy (version v1.6) that applies to every platform regardless of its category. See Taxonomy for definitions, primary and secondary classifications, and the limits of the published version history.

Glossary

Feature: One separately available capability of a platform, as the vendor ships it: its own entry in the release notes, with its own availability. A feature has one or more dated states. "Features recorded" counts features; "changes recorded" counts states.
Significance: Major: a new capability or channel, or a material change in how the platform is used. Notable: a meaningful extension of an existing capability. Minor: UI tweaks, small improvements, SDK and connector maintenance. It is assigned when a feature is recorded, by the same rule for every platform, and describes the change, not its quality.
Signal: A company event that is not a product feature: partnership, acquisition, funding, leadership change, rebrand or packaging change.
Under review: An entry flagged as uncertain by the extraction, a pipeline rule or the owner (for example two labels on the vendor page disagree, or two sources give different dates). See what is checked. It is published with its source but left out of counts in the Release Matrix and side-by-side views until checked.
Empty cell / no release recorded: No change was recorded for that platform and capability in the period. It does not mean the platform lacks the capability: the record covers releases since January 2026, not everything a platform can do.
Counts: Counts show recorded activity: what the vendor documented. Vendors publish at different rhythms and levels of detail, and a feature can be classified in more than one capability, so a higher count is not a better, more innovative or more invested platform. Views across platforms (Release Matrix, Overview, side-by-side views, Insights) therefore show states, or shares within each platform's own record, never totals side by side, and leave minor updates out by default. Counts appear only where they describe the evidence for one platform, such as the changes behind a single Matrix cell. Coverage per platform (sources, date precision, delay) is on the Status page.

Scope

The record covers 15 platforms from January 2026. Partnerships, acquisitions, funding, leadership changes and rebrands are recorded as signals, not features. Bug fixes and documentation-only changes are not recorded.

Which platforms are included

Category. An enterprise platform whose core is customer engagement (CEP), customer data (CDP) or B2B marketing automation (MAP B2B). A product that spans categories is listed under its main one; its features are classified by capability, whatever the category.
Readable public sources. The vendor publishes dated release notes, a changelog or equivalent product documentation that can be read automatically. Platforms whose only public sources are press releases, SDK changelogs or notes published months after the fact are on hold until such a source exists, so that every platform is recorded with the same depth.
Product boundaries. Each platform is the product as the vendor sells it to marketers. Adobe Experience Platform is tracked as a whole (Real-Time CDP included). Salesforce Marketing Cloud is split into Engagement and Next, which have separate release notes. HubSpot covers Marketing Hub, its AI features and the platform features a marketer uses; items available only in other hubs are left out. Chat and events or webinars are in scope for every platform, whichever module ships them.
No endorsement. Inclusion is not a recommendation and absence is not a judgement. No vendor asked or paid to be included.
Adding a platform. New platforms are added when they meet these criteria, with the same rules and, where their sources allow, history back to 1 January 2026. Sources are listed on Sources from the first read.

Corrections

Every feature page and every market signal has a Report an issue form. Reports are checked against the public sources permitted by the source policy: see how corrections work.

Feature pages distinguish the vendor's dated lifecycle states from record acquisition and update dates. An update timestamp does not describe the correction made or establish that a human reviewed the record. Every change to a published fact is listed in the corrections register and on the record; technical rebuilds (a platform re-read from scratch, a product split) and new data-model fields are not corrections and are not listed there.

Neutrality

No vendor pays to be included or reviewed. Wording is original; nothing is copied from vendor text beyond names.

The author implements and is certified on several tracked platforms: see the disclosure. Links to his analysis are hand-picked, at most one per platform or capability page, dated and labelled as opinion.