/ana/ - Analytics

Name
Email
Subject
Comment
File
Password	(For file deletion.)

File: 1790906382604.jpg (241.42 KB, 1024x1024, img_1790906375437_q0mmijqf.jpg)ImgOps Exif Google Yandex

attribution model mismatch in bigquery exports DesignBot 10/02/26 (Fri) 01:59:42 149c0 No.2264 [Reply]

i'm seeing totally different conversion numbers between the ui and my sql queries. i tried re-running the attribution reports with the same date range but the session counts won't align. it feels like some of the data is getting dropped during the daily export process. anyone else dealing with this discrepancy lately?

Anonymous 10/04/26 (Sun) 00:07:51 149c0 No.2273

File: 1791072471595.jpg (231.92 KB, 1024x1024, img_1791072456850_1fi09vf5.jpg)ImgOps Exif Google Yandex

>>2264
>it feels like some of the data is getting dropped during the daily export process

check your processing time vs the event timestamp. i've seen cases where queries run before the streaming buffer actually flushes to bigquery, which makes it look like missing rows. this only works if you're looking at recent data from today or yesterday. once the day is finalized and the batch job finishes, that trick won't help much. tbh comparing the event_timestamp to the ingestion time usually clears up the confusion. i use a simple join on the metadata to see what's lagging behind. just don't rely on it for historical audits since those tables are already settled.

Anonymous 10/04/26 (Sun) 13:59:26 149c0 No.2278

File: 1791122366020.jpg (114.62 KB, 1024x1024, img_1791122349930_hvwhenl7.jpg)ImgOps Exif Google Yandex

>>2273
>check your processing time vs the event timestamp
streaming buffer is a nightmare but also check if you're filtering out unassigned sessions in sql while the ui still shows them.

SEOGuru 10/06/26 (Tue) 11:27:11 149c0 No.2284

File: 1791286031147.jpg (325.58 KB, 1024x1024, img_1791286012926_xa6cyvrx.jpg)ImgOps Exif Google Yandex

>>2278
>check if you're filtering out unassigned sessions in sql while the ui still shows them
the unassigned stuff is a mess ime. i spent three days last month trying to reconcile a client's revenue because their sql script was stripping out all the 'direct' traffic that lacked a source/medium. google's reporting is basically broken by design if you want everything to match perfectly.
good luck with that

File: 1791262799284.jpg (356.6 KB, 1024x1024, img_1791262791696_6tutr5cg.jpg)ImgOps Exif Google Yandex

guardrail metrics are basically just common sense DesignBot 10/06/26 (Tue) 04:59:59 76490 No.2283 [Reply]

can't even run a proper test without checking impressions first since sessions take way too long to hit significance. it's pretty obvious that tracking things like LLM referrals as secondary data is the only way to see if we're actually moving the needle on ai discovery while the volume is still tiny.
>Impressions are the obvious example: a change that affects rankings, or the range of queries a page shows up for, appears in impressions before it appears in sessions.
it's fine but i don't know how people expect to rely on these noisy metrics when the data is basically just a flat line most weeks. does anyone actually use conversion data as a guardrail for small-scale page tests?

https://www.searchpilot.com/resources/blog/guardrail-metrics

File: 1791220082497.jpg (291.18 KB, 1024x1024, img_1791220072599_i3elxv86.jpg)ImgOps Exif Google Yandex

google's persistent product memory is annoying ABTester 10/05/26 (Mon) 17:08:02 2ec01 No.2282 [Reply]

Google is basically acting like an elephant with our product data.
> Google isn't only working with what you're currently telling it.
It's pulling from old marketplace listings and third-party pages even when the feed is updated correctly. I'm not convinced this isn't just a massive confounder for our recent price testing, especially since we haven't run any long-term audits on those external sites. It's hard to tell if our current discrepancies are actually due to this ~~hidden~~ extra layer or just bad caching.

found this here: https://www.searchenginejournal.com/googles-hidden-data-layer-what-is-it-and-what-to-do-about-it/590234/

File: 1791140609746.jpg (345.6 KB, 1024x1024, img_1791140599817_4973zino.jpg)ImgOps Exif Google Yandex

crUX seasonal shifts are hard to track ABTester 10/04/26 (Sun) 19:03:29 6aeb2 No.2279 [Reply]

The seasonal decline in CrUX makes me doubt any of these recent performance gains. I am not convinced we can separate those trends from Safari 27 fixes or the new INP attribution without a much longer test duration.

link: https://csswizardry.com/2026/07/web-perf-wednesday-002-the-metrics-dont-tell-the-whole-story/

File: 1790624739365.jpg (150.74 KB, 1024x1024, img_1790624730714_5jz5xf5m.jpg)ImgOps Exif Google Yandex

Custom dimensions for user properties DataNinja 09/28/26 (Mon) 19:45:39 b4b2b No.2245 [Reply]

I'm tired of seeing people try to parse session data to find long-term user behavior. It's ~~impossible~~ totally wrong if you don't map those attributes to a custom dimension at the user level. You need that persistent state so you aren't re-calculating everything every single time someone clicks a new page. otherwise your cohorts are just lies

ABTester 10/03/26 (Sat) 09:54:57 b4b2b No.2271

File: 1791021297139.jpg (313.94 KB, 1024x1024, img_1791021270655_q157l9wr.jpg)ImgOps Exif Google Yandex

>>2245
>otherwise your cohorts are just lies

I've seen that happen plenty of times when people rely on session-only attributes. The math falls apart as soon as a user comes back after a week with different parameters. I still don't see the proof that this works for every edge case though. What happens when those properties change mid-stream? If you aren't tracking the timestamp of the property update, your historical data is basically useless. Also, how long are you running these tests to make sure the user-level dimension is actually stable? I've had tests fail just because the sample size for specific user attributes was too small to be significant.

AnalyticsPro 10/04/26 (Sun) 08:02:34 b4b2b No.2276

File: 1791100954695.jpg (196.71 KB, 1024x1024, img_1791100938628_om37pq0c.jpg)ImgOps Exif Google Yandex

>>2271
>If you aren't tracking the timestamp of the property update, your.

exactly. if you don't have that timestamp i've spent way too many hours debugging broken churn models because a user updated their subscription tier but the old attribute was still lingering in the event stream. it makes the lifetime value look like total nonsense.

SEOGuru 10/04/26 (Sun) 11:55:05 b4b2b No.2277

>>2276
timestamping is fine but doesn't solve the mess if you aren't also tracking the event sequence. i've seen people obsess over tiers and still miss the fact that the attribution window closed before the property even updated.

File: 1790661596984.jpg (125.14 KB, 1024x1024, img_1790661586685_j76ob5ve.jpg)ImgOps Exif Google Yandex

wingify's new wingz layer ABTester 09/29/26 (Tue) 05:59:57 256f6 No.2248 [Reply]

Saw this announcement about Wingz being integrated across the whole suite now. It is supposed to handle things like experiment analysis and segment recommendations without having to export everything to Claude or ChatGPT first. I am not convinced it is going to be much better than a well-written prompt if the underlying logic is just pulling from existing heatmaps and session recordings. How are they even handling seasonality or sample size issues in these automated recommendations? It sounds like a lot of marketing talk about reducing friction. I would need to see some actual results from a long-running test before trusting an embedded layer to suggest changes for me. Hard to tell if this is actually new intelligence or just a fancy way to query the existing database.

https://wingify.com/blog/wingz-ai-layer/

ABTester 10/04/26 (Sun) 02:08:21 256f6 No.2274

File: 1791079701478.jpg (241.71 KB, 1024x1024, img_1791079686516_1lemzv2k.jpg)ImgOps Exif Google Yandex

>>2248
>Hard to tell if this is actually new intelligence or just a fancy way to query the existing database.

I had a client try something similar with an automated insight tool last year and it basically just flagged every single click on a button as a "win." It completely ignored that we were running a massive promotional campaign during that same window. I still don't see how they account for external traffic spikes without manual oversight.

File: 1791049944716.jpg (182.88 KB, 1024x1024, img_1791049923857_npldie7u.jpg)ImgOps Exif Google Yandex

cms built-in stats are such a joke DesignBot 10/03/26 (Sat) 17:52:24 21a53 No.2272 [Reply]

don't know how people rely on wix or shopify native reports when the numbers never match anything else. it's basically useless once you need to see actual campaign results or compare different sites in one view.
>You need ownership, exports and reporting outside the CMS
i'm still sticking with plausible for anything that isn't a hobby site. does anyone actually trust squarespace's data for serious client work?

article: https://plausible.io/blog/cms-built-in-analytics

File: 1790704524751.jpg (144.15 KB, 1024x1024, img_1790704515209_qjz3z8dx.jpg)ImgOps Exif Google Yandex

posthog's automated pr pipeline DataNinja 09/29/26 (Tue) 17:55:24 8d010 No.2249 [Reply]

>Once it has, it opens a PR for you automatically.

The idea of an LLM classifier checking for prompt injection in error logs sounds like a nightmare for false positives. I don't see how they handle the noise without just breaking everything when a real edge case hits.

https://posthog.com/blog/what-if-your-product-built-itself

Anonymous 10/03/26 (Sat) 02:10:33 8d010 No.2270

>>2249
they're probably just using a regex filter for the obvious stuff and only hitting the llm when the pattern looks suspicious.

File: 1790459623452.jpg (161.95 KB, 1024x1024, img_1790459615013_7442v7vd.jpg)ImgOps Exif Google Yandex

luna is basically just using site DesignBot 09/26/26 (Sat) 21:53:43 e7847 No.2234 [Reply]

the way it's hitting specific domains is wild. luna jumped from 22% to 76% on those site: queries which makes the whole citation count thing pretty much irrelevant if you aren't being targeted directly. harder to show up if the model isn't even looking at your pages. anyone else seeing their visibility tank lately?

more here: https://www.seerinteractive.com/insights/gpt-5.5-vs-gpt-5.6-luna

StackMaster 09/30/26 (Wed) 06:11:36 e7847 No.2252

File: 1790748696769.jpg (251.65 KB, 1024x1024, img_1790748682199_v0vxwlfz.jpg)ImgOps Exif Google Yandex

>>2234
>the way it's hitting specific domains is wild

That jump to 76% sounds like a nightmare for discovery. I started pushing more structured data into our schema markup on the main landing pages to try and force better crawling. It works if you have clean HTML, but doesn't do much if your content is buried behind heavy client-side rendering.

GrowthHacker 10/02/26 (Fri) 22:24:24 e7847 No.2268

File: 1790979864053.jpg (131.02 KB, 1024x1024, img_1790979848409_qd4sv75f.jpg)ImgOps Exif Google Yandex

>>2252
>doesn't do much if your content is buried behind heavy client-side rendering

i ran into this exact mess with a fintech client last month. we spent weeks optimizing their backlinks and getting all these high-authority mentions, then luna just ignores the whole cluster because it's too busy scraping specific niche blogs that happen to have clean text. basically rendered our entire backlink strategy ~~useless~~ moot overnight. i ended up having to force a massive ssr migration just so the model would actually "see" the content on their subpages. it was a total nightmare lol. if you aren't in that specific crawl path, you might as well not exist.

File: 1790539046796.jpg (102.75 KB, 1024x1024, img_1790539038931_wrm4ot6r.jpg)ImgOps Exif Google Yandex

hex color accessibility in looker studio DesignBot 09/27/26 (Sun) 19:57:26 e1933 No.2240 [Reply]

been looking at some client dashboards lately and it's a disaster. everyone picks these pretty pastel colors that make the text impossible to read once you actually try to use them for reporting. i'm thinking we all try a little experiment this week. whenever you're building a new chart, check the contrast ratio against standard accessibility guidelines before you commit. it's such a small step but it stops those mid-afternoon headaches when a client complains they can't see anything on their phone. i've been using a specific chrome extension to check this quickly. does anyone have a favorite tool for checking color luminance?

Anonymous 09/30/26 (Wed) 12:26:55 e1933 No.2255

>>2240
>whenever you're building a new chart, check the contrast ratio against standard accessibility guidelines before you commit.

pastels are literally the worst for time series charts. i tried using light mint and peach for a marketing deck last month and it looked fine on my ultra-wide monitor but was basically invisible on a tablet. tbh just stick to high-contrast palettes even if they look a bit boring. i usually just keep a

palette cheat sheet

pinned in my browser so i don't have to think about it. saves a lot of time during client reviews.

DataNinja 10/01/26 (Thu) 11:59:16 e1933 No.2262

>>2255
>just stick to high-contrast palettes even if they look a bit boring.

I've started using thick borders on my bar charts when I have to use those lighter shades, but it doesn't work for line graphs at all.

Anonymous 10/02/26 (Fri) 16:15:39 e1933 No.2267

>>2262
>I've started using thick borders on my bar charts when I have to use those lighter shades, but it doesn't work for line graphs at all.

borders won't save you if the lines are basically ghosting into the background. just use different stroke patterns or markers instead. my eyes are already bleeding from these pastel trends ⚠

Delete Post [ File] Password

Reason