Your instinct to pull log files is correct. That 15-20% range you mentioned isn't random noise; it often correlates directly with session-level aggregation thresholds. While user1402 pointed out multi-device sessions, I've found the primary driver is often Ads counting separate eligible auctions within a single user session for repeat queries, while Search Console dedupes to a single impression for that query-session pair.
My audit for a B2B SaaS client showed a 22% delta that was almost entirely attributable to logged-in users performing the same navigational search multiple times in a single work session. Ads counted each as a discrete auction opportunity. Your log analysis should key on session IDs and query repetition. The variance won't be a flat percentage; it'll fluctuate with user intent patterns, being highest for informational keywords and lowest for direct commercial ones.
If you're buying media based on auction opportunities, the Ads number is technically "correct" for that purpose. But if you're measuring genuine user exposure for brand lift or share-of-voice, the Search Console figure is closer to the truth. The audit isn't to find which is wrong, but to establish the expected session-based inflation factor for your specific vertical, which you can then use to weight your performance models.
Measure everything, trust only data
That's a big difference, and I'm not surprised you're going to the logs.
When you say you're looking at raw log file data, are you planning to compare session boundaries specifically? I'm trying to understand how to even start that kind of audit.
If Ads counts each new query in a session and Search Console doesn't, how do you decide which model is actually correct for your buying decision?