AI Overview Optimization: What Actually Moves the Needle

Most “how to rank in AI Overviews” writeups are vibes. We wanted controlled tests.

For three months we ran a structured experiment across two retainer client estates: pick 120 commercial queries where a client URL already had some visibility (positions 4–20 in classic SERP, intermittent AI Overview citation), then apply one of 40 on-page changes per URL/query pair and measure AI Overview citation rate before and after. Same query mix, same measurement windows, same engine.

What moved the needle, what didn’t, and what we now ship first.

How we ran it

Each test was a single change applied to a single URL targeting a single query. The change-categories included:

• Title and meta rewrites
• H1/H2 restructuring
• FAQ block additions and rewrites
• Schema additions (FAQ, HowTo, Article, Product variants)
• Date and recency signals
• Internal link additions from authority hubs
• Outbound citation additions to authoritative sources
• Original-data block additions (small tables, charts)
• Brand-mention removals (yes, removals — see below)

We measured citation rate as: “of the next 20 daily checks at the same time of day, in how many did the URL appear in Google’s AI Overview for the target query?” before-and-after, with a 14-day wash between phases to let the index catch up.

We didn’t randomize across all 4,800 possible cells. We applied changes where the existing baseline gave the change room to move. That’s a real limitation — these results generalize to mid-funnel commercial queries with some prior visibility, not to cold-start queries.

What worked

Ranked by mean change in citation rate (percentage points, change minus baseline):

1. Adding original-data blocks — small tables or charts with proprietary numbers. +38pp mean lift across the 7 tests in this category. The lift was concentrated in informational and commercial-research queries.
2. FAQ block with real questions — questions taken from People Also Ask, Reddit, or sales-team logs, with concise direct answers. +24pp. Same FAQ block written with generic “What is X?” questions: +6pp. The question quality mattered more than the schema markup.
3. Recency signals in H1 and intro — explicit “Updated [current month]” in the H1 area, with a refreshed intro that referenced a recent industry event. +19pp. Updating the publication date in JSON-LD only (no visible recency in the page): +4pp. Visible > markup-only.
4. Authoritative outbound citations — 2–4 outbound links to clearly-authoritative sources within the body. +14pp. The lift was larger when the linked sources were also commonly cited by AI Overviews for the same query.
5. Internal links from authority hubs — adding 2–3 internal links from the site’s top pages to the target URL. +11pp. Lift was higher on sites with a clearer hub-and-spoke structure.

What didn’t work

Same magnitude (percentage point change), with effects we’d called negligible:

• Adding Article schema to pages that didn’t already have it: +3pp. Statistically indistinguishable from no change in our sample.
• Title rewrites optimized for the exact query: +5pp. Helped classic ranking, didn’t notably help AI Overview citation.
• Meta description rewrites: +1pp. AI Overviews don’t appear to read these.
• Removing brand mentions (we tested this to see if dense brand language was crowding out citation triggers): −2pp. Don’t bother.
• Increasing content length by ~40% with related material: +4pp. Adding length for its own sake doesn’t move citation share.

The two things we expected to work that didn’t were schema-only changes and pure title optimization. Both have been the loudest “AI Overview hacks” of the past year. Neither holds up in this dataset.

Combination effects

We tested seven combinations of two changes (the highest-leverage pairs). The most striking:

• Original-data block + FAQ with real questions: +51pp. Sub-additive (the two effects don’t fully stack — there’s an implicit cap because citation rate is bounded), but still the strongest single intervention we measured.
• Recency signals + authoritative outbound citations: +28pp. Cleanly additive.
• Internal links + FAQ: +30pp. Slightly super-additive — internal links appear to amplify the FAQ effect by getting the page crawled and indexed faster.

What we ship first now

When we onboard a new retainer and the goal is AI Overview share for mid-funnel queries, our default first-90-days move is:

1. Identify 5–10 top target queries with existing baseline visibility (positions 4–20 in classic SERP).
2. For each, add a proprietary-data block at the top of the target page (table or chart, even a small one).
3. Add a FAQ block with 4–6 real questions sourced from sales logs and PAA.
4. Add 2–3 internal links from the site’s strongest pages.
5. Push a visible recency update in the H1 area and intro paragraph.

That’s it. No schema heroics, no title tortures, no length stuffing. We’ve now run this pattern on ~30 retainer setups since closing this study; the per-query citation-rate lift averages +34pp in the first 8 weeks. The variance is wide — some queries didn’t move at all — but the median is high enough to make this the default playbook.

What we’d test next

Two follow-ons we’re queuing:

1. The “original-data block” effect across non-Google AI engines — does adding small proprietary tables also lift citation share in Perplexity, Claude, and ChatGPT? Early signal: yes for Perplexity and Claude, neutral for ChatGPT.
2. Decay timing — how long do these lifts last? Anecdotally the AI Overview lift starts decaying around week 10–12 if there are no further updates. We’re queuing a structured 26-week follow-up to measure the decay curve directly.

If you’re running AI Overview work in 2026 and your stack is full of “we should add schema” experiments — try the data-block + real-FAQ + internal-links combination first. It’s the only thing in our dataset that consistently moved.