AI Share‑of‑Voice Case Study: 40% Gain in 90 Days (2025)

If AI answers increasingly shape how people research and evaluate brands, how do you prove your clients are actually showing up in those answers—and increase that presence quarter over quarter? This case study walks through the exact workflow our agency used to lift an anonymized client’s AI Share‑of‑Voice (SoV) by 40% in 90 days across Google AI Overviews, chat‑style assistants, and Perplexity‑type answer engines.

Disclosure: This is a composite case built from multiple client engagements to protect confidentiality. All methods and numbers reflect real‑world patterns and repeatable agency practices.

What we measured (and why it matters)

AI Share‑of‑Voice is the percentage of answers—across a defined prompt set and time window—where your brand appears, either as a cited source or a recommended entity. Practitioners commonly track both citation‑based SoV (how often your URLs are linked) and entity‑based SoV (how often your brand is named or recommended), then roll results up by engine and topic cluster.

Recent definitions align on this approach. For example, Single Grain frames AI SoV as the percentage of answers featuring your brand over a defined set and period, often weighted by prominence, while urging engine‑ and competitor‑level comparisons. Cassie Clark recommends calculating SoV as brand citations divided by total citations with multi‑run averaging and maintaining a separate entity score for recommendation prompts.

Why focus here now? Consumer adoption of AI search is material. In June 2025, the Pew Research Center reported that 34% of U.S. adults have used ChatGPT, with 58% of adults under 30 reporting usage, and their browsing‑data analysis found that about six‑in‑ten users encountered AI summaries in search results pages; see Pew’s short read 34% of U.S. adults have used ChatGPT (2025) and data‑labs analysis What web‑browsing data tells us about how AI appears online (2025). McKinsey’s October analysis suggests roughly half of consumers already use AI‑powered search, with a projected impact of up to $750B in U.S. revenue by 2028; see New front door to the internet: Winning in the age of AI search (McKinsey, 2025).

Two measurement considerations guided our approach: volatility by engine and query intent, especially within Google’s AI Overviews (AIO), and cross‑engine differences. BrightEdge’s one‑year review documents heavy keyword churn and fluctuating coverage by category; see the AIO Overviews One‑Year Review Research Paper (BrightEdge, May 2025). Trade reporting later in 2025 noted that AIO coverage has pulled back for purchase‑intent queries while remaining active for research; see Google AI Overviews retreat for sales queries but guide research (Search Engine Land, Nov 2025). Cross‑engine traits also matter: studies in 2025 observed that Perplexity‑style engines often cite third‑party and UGC sources at higher rates, while ChatGPT‑style assistants tend to mention brands more frequently than they link them; see AI search engines often cite third‑party content (Search Engine Journal, Feb 2025) and Chat assistants recommend brands more than they cite them (Search Engine Land, Aug 2025).

The 90‑day plan at a glance

We ran a five‑phase loop:

• Audit: Establish a 120‑prompt set across research and recommendation intents; define engines; set baselines for entity and citation SoV; flag gaps by topic and engine.
• Optimize: Strengthen content and technical signals (schema, entities, author pages), create missing assets, and tune format for likely citation.
• Amplify: Digital PR and expert distribution to seed third‑party coverage and UGC likely to be cited by answer engines.
• Monitor: Weekly checks by engine/topic; hallucination triage; prominence tracking; adjust tactics.
• Iterate: Publish incremental updates, recycle PR angles, refine prompts, and recalibrate the prompt set monthly.

Owners were split across SEO, content, PR, and analytics. Cadences: weekly working reviews, monthly executive roll‑ups.

Day 0 baseline and gap analysis

The composite client (B2B SaaS, mid‑market) operated in a competitive category with strong incumbent brands. We defined a 120‑prompt set: 70 research prompts (e.g., “how to evaluate [category] platforms,” “pros and cons of [solutions]”) and 50 recommendation prompts (“best [category] tools for [use case]”). Engines were grouped as Google AI Overviews, Perplexity‑style answer engines, and chat‑style assistants. We averaged results over three runs for chat engines to account for variability.

At baseline, entity SoV was present but inconsistent in recommendation prompts—the brand surfaced as a secondary mention in about a quarter of assistant‑style answers. Citation SoV lagged in research prompts on Perplexity‑style engines, where third‑party roundups dominated. AIO citations appeared sporadically, leaning toward a small set of evergreen guides.

Below is a compact view of baseline vs. Day 90. Percentages represent rolled‑up SoV across the 120‑prompt set, averaged over a 14‑day window.

Note: Composite numbers; variation by topic cluster ranged ±7–10 points.

Optimization moves that moved the needle

Google AI Overviews

We strengthened structured data and authorship signals on our highest‑utility guides—Article and How‑To schema, consistent author bios with credentials, organization and product schema on relevant pages, and tighter internal linking from topic hubs to supporting pages. The goal was eligibility and clarity for snippet‑style grounding, in line with Google’s public guidance on appearing in AI features and succeeding in AI search; see Succeeding in AI search (Google Search Central, May 2025). We also consolidated competing URLs to reduce topic dilution, refreshed aging content with new primary research, and embedded expert quotes to elevate E‑E‑A‑T signals.

Perplexity‑style engines

We built and pitched third‑party explainers and comparison pieces to credible publications and community sites (industry newsletters, analyst roundups, community review hubs, YouTube explainers). Reporting in early 2025 indicated these engines often cite third‑party and UGC sources heavily; we leaned into that pattern by supplying quotable artifacts and clean, sourceable visuals. We also encouraged product‑agnostic tutorials from partner creators and customer advocates, timed with feature releases, to diversify non‑owned citations.

Chat‑style assistants

We increased brand authority signals: executive‑bylined thought leadership with unique data; consistent naming conventions across properties; and FAQ‑style content that directly answers common evaluator questions. Industry commentary across 2025 suggested assistants tend to mention brands more frequently than they link them; we prioritized entity clarity and reputation. We also published a short “why us for [use case]” series mapped to recommendation intents, each with transparent pros/cons and comparison criteria that assistants could summarize coherently.

Digital PR and distribution

We shifted PR from pure backlink acquisition to authority and mention strategy: social‑first video snippets for journalists and creators, fast reactive data stories, and multimedia press assets to raise pickup probability. Late‑2024/2025 guidance from leading PR and marketing outlets emphasized that such tactics improve trusted mentions and citation potential. To operationalize this, we established a monthly “evidence calendar” tying product news, data stories, and community events to pitch angles; distribution spanned LinkedIn, YouTube, and niche forums.

Risk and quality controls

To keep quality high and reduce noise, we added a lightweight hallucination triage. When assistants misattributed features or invented comparisons, we published clarifying pages, tightened product names and descriptions, and updated FAQs. We also monitored for prompt manipulation or poisoning attempts and followed prudent controls inspired by OWASP’s LLM prompt‑injection prevention guidance (2025), including clear delimiters in demos, provenance notes on datasets, and post‑publication validation.

Monitoring and iteration

Weekly, we reviewed SoV by engine and topic cluster, scanning for trend breaks and volatility. Google AIO required special handling: given reported coverage fluctuations through 2025, we used rolling 14‑day windows for stability and tracked presence in research‑intent prompts where AIO remained more active.

On tooling, we centralized engine‑level SoV, mentions, total citations, and platform breakdowns in a white‑label dashboard so client stakeholders could see progress without screenshots. An agency‑focused platform like Geneo (Agency) is one option for this, as it monitors chat‑style assistants, Perplexity‑style engines, and Google AI Overviews and aggregates signals such as Share of Voice, AI Mentions, and Total Citations into a trendable view. Comparable stacks can be built with a mix of visibility trackers and custom reporting, provided they support time‑windowed comparisons and prompt‑level detail.

Iteration decisions followed a simple rule: double down where SoV rose (publish adjacent assets, pitch follow‑ups to the same publications) and remediate where it fell (improve entity clarity, refresh or consolidate content, seek third‑party coverage in that exact topic).

Results after 90 days

Aggregate AI SoV improved by roughly 40% relative to baseline. The largest absolute gains came from entity SoV in chat‑style assistants (+15 points) and citation SoV in Perplexity‑style engines (+10 points). Google AIO gains were steadier but meaningful after content consolidation and structured data clean‑up, particularly in research‑intent prompts.

Two ancillary signals trended positively. Branded search impressions and direct traffic rose in parallel with entity SoV gains in assistants, suggesting higher brand recall—though we’re cautious about asserting causality given other marketing in flight. Sales conversations also referenced third‑party explainers published during the campaign, indicating that PR‑seeded assets influenced discovery.

Caveats and confounders

• Volatility is real. AIO presence varies by category and query intent; we saw week‑to‑week swings that required smoothing windows to interpret.
• Assistants can change behavior with model updates; periodic prompt re‑baselining is essential.
• Correlating SoV to pipeline requires careful multi‑touch analysis; we treated SoV as a leading indicator.

Troubleshooting and limits: what nearly derailed us

• Thin or fragmented content: Topic dilution hurt AIO eligibility. Consolidation and clear topical hubs solved it.
• Lack of third‑party proof: Without credible non‑owned sources, Perplexity‑style engines ignored us. We invested in data‑driven PR and community explainers.
• Entity confusion: Inconsistent product naming and author profiles depressed assistant mentions. We standardized names and strengthened author pages.
• Over‑indexing on one engine: Early wins in assistants masked weak Perplexity citations. The engine‑split view prevented a false sense of progress.
• Measurement drift: Prompt sets get stale. We refreshed monthly and versioned prompt lists to keep comparisons fair.

Limitations of this study

• Composite methodology: To preserve anonymity, we combined multiple client patterns; while representative, single‑brand variability can be higher.
• No universal SoV formula: Weighting schemes differ (e.g., prominence scoring). We used an unweighted roll‑up for clarity here and encourage teams to test prominence weights.
• External factors: PR cycles, product launches, and competitor moves can shift SoV independent of our optimizations.

Next steps for agencies

• Define your prompt set by intent and engine; run a two‑week baseline with multi‑run averaging for assistants.
• Fix content eligibility and entity clarity first; then pursue third‑party explainers to earn citations.
• Set a weekly review and a monthly iteration cadence; smooth results with a rolling 14‑day window.

If you need a client‑ready, white‑label view that tracks SoV across AI Overviews, Perplexity‑style engines, and chat assistants with daily history and exportable dashboards, consider an agency platform like Geneo or assemble an equivalent stack from your existing tools.

—

Want to compare your own baseline to this 90‑day curve? Start with 100–150 prompts, split by research and recommendation intents, and ask: where does your brand show up today—and where should it show up next?