Vincze Norbert — 3-month Google Ads edit-history deep dive

Bottom line: the earlier audit was wrong by more than an order of magnitude — and “he did almost nothing” does not survive the numbers.

First, the number Matt was working from is wrong. The 2026-07-24 audit reported 20 Norbert change events over the accounts' entire lifetime. Queried directly against the Google Ads API, Norbert made 425 change events in 30 days — the clear majority of all change activity across the eight accounts in that period, the remainder being Google's own Auto-Apply and Matt's API calls. That earlier figure was wrong by more than an order of magnitude and no conclusion should rest on it.

Second, the 30-day number alone would have produced a wrong answer. Google's API attributes changes to a named user for 30 days only (2026-06-27 → 2026-07-25). A second API resource, change_status, reaches 90 days back but records no user: it shows 6309 resources changed across 6 of the 8 accounts between 2026-04-28 and 2026-06-26, proving the earlier period was busy without saying whose hands did it. To close that gap, the Google Ads UI change history — the only source that names users further back — was scraped directly from the logged-in browser for the accounts where it mattered most, adding 253 further change-history entries covering 1583 individual items on 6 accounts. Where that was done, the picture changed materially. Where it was not, the honest reading of an idle gap is “no attributed activity”, not “nothing happened” — the coverage table below says exactly which is which.

Two things the UI history settled that the API could not. (1) The single densest burst of activity in the whole quarter — 2 529 resources changed on 2026-06-25 on Csupi, two days outside the API's attribution floor — was Norbert, and nobody else: 19 change-history rows / 32 individual changes between 19:11:42 and 19:25:48 that evening, mixing a Google Ads Editor bulk upload of 7 responsive search ads with 17 budget changes made by hand across the geo campaigns. (2) Fusion Tables was not abandoned. The API window shows 2 events there and an earlier draft of this report called it “effectively unmanaged”; the UI history shows a full account build in May — 60 entries, 403 items, 8 working days, including 76 negative keywords and 9 responsive search ads uploaded through Editor.

And there is a pattern in the corrections themselves that Matt should weigh. Every account for which the full UI change history was pulled moved up this table, not down. Fusion Tables went from “2 events, effectively unmanaged” to a full account build. Komplex Logopédia went from 3 working days to 9, with the tightest cadence of any account (never more than 26 days between sessions). Lovasember gained a campaign build. Dömösi gained three working days and its real conversion setup. Not one account looked worse under closer inspection. The lesson for the earlier audit is blunt: the 30-day API window is not a sample of the quarter, it is a keyhole, and judging anyone through it produces the wrong answer.

The verdict on everything we can attribute: real work, delivered in bursts. Across both sources, Norbert's 678 attributed changes fall on 30 distinct calendar days in the quarter, and 393 of them (58%) landed on just four days: 2026-07-16, 2026-06-25, 2026-07-06, 2026-07-01. That is the shape of the work — a handful of dense, competent sessions with long quiet stretches between them, rather than steady week-by-week management.

The single clearest answer to “what did I actually pay for?” is a date range. Between 2026-05-22 and 2026-06-23 — 32 days, a third of the quarter — there were 2 changes in the entire portfolio, across all eight accounts, while roughly half a million forints of client budget kept spending. That is the gap, and it is now provable rather than inferred: the UI change history covers that period in full for six of the eight accounts, so this is real silence, not missing data.

The quality of what was done is mostly good. The negative-keyword mining is well targeted (competitor names, wrong districts, irrelevant intents — not padding), the Kertújítók geo-targeting rebuild from business-radius to explicit proximity radii is skilled work, and Dömösi's two July sessions coincide with a genuine jump in demand signals. The tooling is consistent with real account work rather than activity theatre: hand edits in the web UI for the fine-grained tuning, and Google Ads Editor for the bulk builds (the Fusion Tables May build, the 2026-06-25 Csupi session). Editor is the tool that under-logs, so those bursts represent more work than their entry counts suggest, not less.

But it is periodic intervention, not ongoing management. Across three months and eight accounts carrying 1 573 778 HUF of client ad spend, the attributed evidence shows roughly a dozen genuine working sessions. There was not one bid-strategy change on any account in the entire window, and conversion tracking was touched on only three of eight accounts in three months (Fusion Tables in May, Dömösi on 06-24, Kertújítók on 06-25) — on accounts where conversion efficiency was visibly declining (Komplex: 10 conversions in May, 4 in June, 4 in July on ~90 000 HUF/month). The work that happened was competent; the work that did not happen is the gap.

The single biggest problem in this portfolio is not in the change history at all — it is what these accounts call a “conversion”. Across the eight accounts in the last 30 days, 596 conversions were recorded. Only 36 of them are a completed action — a purchase, a booking or a submitted form. 110 are taps on a phone number, and 450 are clicks on buttons or views of pages. Dömösi reports 463 and has 11 real bookings. Lovasember's 17 conversions are all appointment_cta_click. Csupi — an emergency towing business — counts only phone-number taps, with no verified call tracking at all. Komplex's 4 are clicks on a booking link. These are the numbers Smart Bidding optimises toward and the numbers any performance review would quote, and they are measuring intent, not outcomes. And this is not neglect — it is a deliberate configuration. The UI change history shows Norbert working on conversion settings on three of the eight accounts. On Kertújítók (2026-06-25) he did exactly the right thing: recategorised the “Ajánlatkérés” and “Küldés” button conversions out of the “Page view” category and added Begin checkout to the account goals. On Dömösi (2026-06-24) he went the other way, spending a session assigning conversion values to the cta_apartman_foglalas_click-type button clicks and rewiring the campaign goals around them. So he knows how this works; on Dömösi he chose to make booking-button clicks the optimisation target. That is a defensible choice for a campsite with no online checkout — but it means the account's headline “463 conversions” must never be read as 463 bookings, and the remaining five accounts were left with whatever they had.

Three things need a decision this week, not next quarter: (1) Komplex Logopédia has been fully switched off since 2026-07-21 — zero enabled campaigns after 255 195 HUF of spend, and the client may not know; (2) Csupi's Recommendations Auto-Apply is deleting keywords faster than they are added (11 removed on 07-16, 11 on 07-17, 11 on 07-21), which is why the identical set of 11 keywords was re-entered by hand three times in six days — one setting fixes it permanently; (3) settle what a “conversion” means across the portfolio before any performance judgement is made — 76% of the portfolio's recorded conversions are button clicks or page views. Fix that first, because every other performance conversation depends on it. That is the largest gap this analysis found, and it is a gap in scope rather than in effort.

One suspicion checked and dismissed. Fourteen “dömös kemping / dömös camping” terms were added as exact negatives on Dömösi Kemping on 2026-07-16, which reads as a manager blocking his own client's brand. Verified against the live account: the client's brand is Danu Kemping, the Brand_danu ad group still targets it, and the blocked terms are generic category+location searches being routed to the sibling “nonbrand” ad group. Correct structure, not sabotage — recorded here because it is exactly the kind of finding that would have been reported as damning if nobody had opened the account.