SEER AI LABS
Do Meta Descriptions Influence AI Visibility? How De-Optimization Impacted GEO
Key takeaway: Across a six-week test, altering meta descriptions had no measurable effect on how often AI crawled or cited our pages.
The Challenge
Google has been clear that meta descriptions aren't a ranking factor, and previous Seer research found that Google is rewriting meta descriptions ~70% of the time. Most teams still write them to improve organic click-through rates, but as meta descriptions are continually deprioritized in traditional search, is there any value in 'optimizing' them for AI search?
Type
What We Tested
Seer has spent years refining how we write meta descriptions - from tactical tips to integrating paid data to write optimized meta descriptions. So instead of reinventing the wheel and spending more time 'optimizing' something that may no longer be impactful, we did the opposite: we sabotaged existing meta descriptions.
Our hypothesis: if meta descriptions are a relevance signal that AI crawlers use to decide what's worth retrieving, then de-optimizing them on strong, high-traffic pages should drop AI crawls and citations on those pages.
Validation: Google already rewrites meta descriptions roughly 70% of the time, so their influence on traditional search is shrinking. The open question is whether AI search treats them any differently.
Strategy
We took 21 of our high-traffic pages and changed their ‘optimized’ meta descriptions into a punctuation mark or test statement:
- On 11 of them, we replaced the description with a single period ("."), effectively removing it
- On the other 10, we swapped in a neutral placeholder: "We intentionally changed this page's meta description for an AI search experiment."
We kept a control group of 10 untouched pages, so we could separate any real effect from a sitewide change in bot traffic. If meta descriptions carried weight for AI, the de-optimized pages would fall behind the untouched control group.
Results
→ Consistent rise in AI bot traffic on all pages: blanked (+22%), placeholder (+12.6%), and control (+21.7%)
→ -1.3% true effect (statistically zero) on AI retrieval crawl once we account for the sitewide trend
→ No AI citation penalty on the de-optimized pages
Breaking the descriptions had no measurable effect on the AI crawlers or citations. Every group (blanked, placeholder, and untouched control) gained AI bot traffic after May 1 at nearly the same rate. Seeing the control group move alongside the treatment group signaled a sitewide trend that wasn't caused by any of our meta description changes.
AI BOT TRAFFIC - WEEKLY - INDEXED
Every group rose together, including pages that were left alone
To confirm, we modeled the treated pages against a Bayesian counterfactual of how they'd have performed if left untouched. The real effect came out to −1.3% (95% CI: −5.9% to +3.3%), a statistical zero. Citations followed the same pattern: no drop, no lift tied to the experiment. Section citations moved with the same wave shape we saw in the months before May 1.
TREATED PAGES - ACTUAL VS. MODELED COUNTERFACTUAL
Actual crawl stayed inside the no-effect band all post-period
Note: The crawl figures above reflect the ChatGPT-User agent, the metric where our control predicts treated-page behavior most reliably. Other AI crawlers were included in the broader analysis and showed the same no-effect pattern.
Next Steps
Meta descriptions are not where to spend GEO or SEO time. Our experiment showed no meaningful change in crawler traffic or citations across treated and control groups. Based on that evidence, we're confident meta descriptions are not a lever for a site's AI visibility.
That doesn't mean you should leave them blank, but for the bulk of your site, focus your time on other signals that benefit both AI search and humans.
About the Author
Bryan Gunawan
AI Optimization Intern
Bryan Gunawan is an AI Optimization Intern at Seer Interactive, where he's running experiments and research focused on how different elements shape AI visibility. Before Seer, Bryan spent two years as a full-stack developer at Launchpad, where he built and deployed 10+ production applications and designed AI-assisted workflows across multiple client projects.