What to fix first when everything needs fixing

When a marketing function is behind on everything, the instinct is to start with the visible things: the website design, the content calendar, the campaign that is underperforming. That order is wrong, and it is wrong for a mechanical reason.

Three things sit upstream of everything else and cap it. Fix them and the same campaigns produce more. Skip them and every improvement downstream is measured through a broken instrument.

First: time to first byte

Core Web Vitals thresholds have not changed. Largest Contentful Paint good under 2.5 seconds, Interaction to Next Paint good under 200 milliseconds, Cumulative Layout Shift good under 0.1, all judged at the 75th percentile of a 28-day rolling window. Anyone telling you Google tightened LCP to 2.0 seconds in 2026 is repeating machine-generated blog filler.

Server response time is upstream of LCP. Every millisecond your server takes before sending the first byte is a millisecond LCP cannot get back, no matter how well the front end is built. Aim to keep it under 800 milliseconds at the 75th percentile.

Fixing it is usually infrastructure, not design: caching, a content delivery network in front of the origin, database queries on the templates your traffic actually hits, and hosting located sensibly for Australian users. It is unglamorous, and it lifts every page at once.

Two measurement notes that will save you an argument this year.

Chrome 151 began stable rollout on 28 July 2026 with soft-navigation and interaction-contentful-paint entry types out of origin trial. Chrome can now detect same-document route changes and start a fresh metric lifecycle per route, so vitals get attributed to the in-app route rather than dumped on the initial load. The web-vitals library added native support in version 6 on 21 July 2026. If your site already runs a virtual-pageview shim alongside it, you will double count, and your reported percentiles can get worse purely because previously invisible in-app sessions are now included. Run both in parallel before cutting over, and stamp the browser version and library version on every beacon.

Second, Google's field dataset only sees Chrome and always will. Firefox 144 added INP in October 2025 and Safari 26.2 shipped LCP and the Event Timing API on 12 December 2025, though Safari's INP still reports unreasonable values at times. Neither plans CLS. Your own real-user data and Google's data will diverge more over time, and that divergence is not an error. Google grades Chrome. Your customers are not only on Chrome.

Second: the consent banner

The consent banner is usually the first interaction every visitor has with your site, and it is treated as a legal checkbox that nobody owns.

It does three kinds of damage.

It is often the largest layout shift on the page, because it injects late and pushes content. It is frequently the worst interaction, because the accept button runs a tag manager sequence before it dismisses, which lands directly in your INP score. And it determines whether any of your marketing data exists at all.

That last one changed on 15 June 2026. In GA4, ad_storage became the sole control over whether cookies and identifiers are shared with linked Google Ads accounts. The Google Signals toggle now only governs signed-in behavioural reporting: demographics, interests, cross-device.

Two failure directions exist: accounts with Signals off but ad_storage granted now collect more than they intend, and accounts relying on Signals while ad_storage defaults to denied lose Ads data quietly.

Audit it in ten minutes. Open DevTools, Network tab, filter for google-analytics.com, then compare the gcs parameter and the gcd parameter on a reject and on an accept. Do not use GA4 Tag Diagnostics for this, it lags 48 to 72 hours. Server-side tagging flows are unaffected by the change itself.

Then fix the mechanics. Reserve the banner's space so it cannot shift layout. Dismiss on click and fire tags after. Make the accept path one clear action rather than a maze, because a banner designed to frustrate produces a low consent rate, and a low consent rate means more of your reported conversions are modelled rather than counted.

Third: conversion signal quality

Both major ad platforms now bid off what you feed them. If the feed is wrong, the machine optimises confidently in the wrong direction, and it does it faster than a human ever could.

Work through this list in an afternoon.

1. 1.List every conversion action. For each, ask whether you would happily pay more to get another one. Anything where the answer is no comes out or gets reclassified.
2. 2.Check whether any of those secondary actions are now influencing bidding. Google's journey-aware bidding lets Search Target CPA campaigns learn from non-biddable goals like calls, form fills and newsletter signups, so observe-only is not what it used to be.
3. 3.Turn on enhanced conversions with hashed first-party data.
4. 4.Import offline conversions from your CRM weekly, with real values. A qualified enquiry and a closed deal should not be worth the same number.
5. 5.Confirm your GA4 data-driven attribution eligibility. It needs at least 400 conversions for the specific action and 20,000 across all actions in the lookback window. Below that you are on last click.

The order, and why

Speed first, because it lifts every page and every channel at once and nothing downstream can recover what it loses. Consent second, because it decides whether you can see anything. Signal third, because it decides where the money goes.

Only then content and campaigns, which now land on a fast site, get measured properly and feed a bidder that knows what a good outcome looks like.

Most of what looks like a marketing problem is an instrument problem, and the instruments are cheaper to fix.