# DEVIATIONS — serp-first-screen-census-2026

1. 2026-09-30: pixel distances were dropped and blocks_before is computed from DOM order (v1 pixel version distorted; rows_v1.jsonl unused).
2. 2026-09-30: the first pass ran in the logged-in Chrome. That browser carries a Google Ads account (a "Your Google Ads summary" panel appears) and its IP is in Batam, Indonesia. Paid-listing counts for the SAME query changed between loads (for example 4 then 0 on one plumber search). That pass is NOT used.
3. 2026-09-30: final pass = brand-new logged-out headless Chrome profile per search, gl=sg, hl=en, uule set to Singapore, one search per query, 8 seconds apart, 40 searches. In this configuration the ad block was empty in all 40 searches, while the logged-in browser showed 3 to 4 paid listings on some of the same queries. Paid listings are therefore NOT measured by this study and no paid-listing figure is published. Only map pack, AI overview, People also ask and their position relative to the first organic link are reported.
4. Location caveat: the request asks for Singapore (gl and uule) but originates from an Indonesian IP; Google may weigh the IP.
5. Validation: test-retest of 12 searches (see retest.json) plus a screenshot check (see notes in the article method).
6. Screenshot check (3 fresh loads: plumber singapore, dentist singapore price, florist singapore): map pack above the first organic result confirmed on "plumber singapore". On "dentist singapore price" the screenshot shows an AI Overview panel still loading ("Thinking a little longer") above the first organic result, while the recorded row for that search flags an AI Overview as present but does not place it above the first organic result. So the AI Overview count is a FLOOR (slow-loading panels can be missed). Reported as "at least".
7. 2026-09-30, after collection: CLAIM resolved from the pre-registered template to the single lead ratio 19 of 20 (plain searches, map pack above first organic). The original pre-registration text is preserved as 01b-signature.PREREG-copy-20260930.md (SHA256 in the .sha256 file). The copy was made after collection and before the amendment, not before collection: the pre-registration file itself was written before any measurement (its timestamp and the run logs show this) but no hash was taken at that time.
8. Test-retest (12 searches, second load some minutes later): identical structure on 9; map pack above the first organic result agreed on 12 of 12; AI Overview above the first organic result agreed on 10 of 12 (one gained, one lost). The AI Overview counts are therefore the least stable numbers in the study.
9. 2026-09-30: METHOD, TOOL_OR_SCRIPT, VALIDATION, LIMITS and REPRODUCE lines in 01b-signature.md were rewritten to describe the protocol actually run (clean logged-out headless Chrome, not the logged-in Chrome of the pre-registration). The original text is in the PREREG copy.
10. 2026-09-30: a column paa_above_organic was added to the public CSV so the People-also-ask counts (6 plain, 0 price) are explicit rather than derived; nothing else changed. After an adversarial review, the article wording was tightened: 'first screen' language replaced by 'top of the results page' (reading order, no pixels), and the study is titled 'Google Results-Page Census 2026'.
11. 2026-09-30, after editor-in-chief review: (a) AI-panel figures are now written as floors ('at least') and 'none recorded' as ceilings ('at most') in the article: one 'none' row (dentist price) had a slow AI panel in a screenshot and one (locksmith price) gained an AI panel on retest. (b) The article no longer draws any SEO-versus-ads conclusion: ads were not measured. (c) It now states that the final run was headless Chrome, logged out, from an Indonesian IP, which may itself explain the empty ad block in all 40 searches; it does not imply Google shows no ads to Singapore searchers. (d) The method sentence now discloses: planned first; set-up differed from the written plan and changed once; plan text fingerprinted only after collection (see 7 and 9); headline ratio and claim wording chosen after the counts. (e) The 1.4% vs 3.0% click-study figure now carries that study's own caveat ('valid as a description', 'invalid as a universal prediction table'). (f) Internal links to the two /blogs/news/ pages were swapped for their /blog/ equivalents (both return 200). Deploy note: the CSV must be live at /data/singrank-google-first-screen-census-2026.csv before publishing, and this file should be published beside it.