What should buyers take from September’s developments?

1. How should we track brand visibility in Google AI Overviews and AI Mode?

Start with Google’s own reporting, then add answer-level research where needed. September followed the worldwide rollout of Google’s generative-AI performance reports. Google’s help documentation describes organic impressions, page performance and country/device breakdowns; it also says a missing report usually means insufficient impressions, with exclusion settings another possible explanation. (support.google.com)

Decision: Use native data to investigate your URLs’ exposure. Do not interpret an unavailable report as proof that assistants never mention the company.

2. Is image-led discovery becoming more measurable?

Google added a relevant reporting capability on 24 September. Its new multimodal search filter covers experiences including Lens, Circle to Search, image uploads and Chrome’s image-search function. The announcement applies to Search performance and generative-AI reporting, with export available. (developers.google.com)

Decision: Consumer-product teams should investigate whether this reveals a meaningful discovery route. B2B software teams should first establish relevance to their own buyers rather than automatically commissioning more visual content.

3. Can an API benchmark stand in for the customer-facing assistant?

Not without validation. A 16 September preprint audited 1,536 product-query responses across consumer interfaces and APIs. It reported average source-domain overlap of 12.0% between ChatGPT’s interface and API, and 14.8% for Gemini. Recommended products also varied across repeated requests. These are preliminary findings from a consumer-product audit, not universal platform constants. (arxiv.org)

Decision: Require suppliers to label API observations explicitly. For an important shortlist question, include a documented check of the consumer experience you actually intend to study.

4. Does being cited cause a brand to be mentioned?

The September evidence supports investigation, not that causal conclusion. A preprint analysed 34,960 unbranded prompt-engine observations from 75 anonymised Aiso projects. Own-domain exposure and brand-related retrieval were associated with mention rates, but the author explicitly describes the fitted model as predictive rather than causal. (arxiv.org)

Decision: A competitor citation analysis can identify useful pages to inspect. It cannot, by itself, justify a claim that acquiring a similar citation will produce the same recommendation.

5. Did earned media regain citation share?

In Meltwater’s September sample, yes. Its report, published on 15 October, put the Earned/News category at 35.6% of citations, up from 29.4% in August. The report also explains why source volume and source share can move differently. One month’s change does not establish a durable trend or a return on PR expenditure. (meltwater.com)

Decision: Ask for exact cited pages and engine-specific changes before buying a digital PR programme on the strength of an aggregate publisher chart.