September 2025: Spam Update Ends, FastSearch Powers AI Citations & ChatGPT Ecommerce

August Spam Update & num=100

Google’s August 2025 Spam Update finished rolling out on 22 September after a 27-day run. This was the first, and possibly only, spam update of the year and it was global in scope. The impact was unusually big for a spam update.

At the same time, Google removed support for the &num=100 query parameter, which most rank trackers have historically relied on to pull the top 100 results for a keyword in one request. This change has broken a lot of third-party tools, scrambling position tracking (including in Search Console, with effects on impressions and CTR due to bot traffic dropping). Tools are being forced to rethink their entire architectures, in some cases resulting in them having to make 10× more requests to reconstruct a top-100 list from Google, pushing up their costs significantly. Many tools have simply started limiting their tracking to the top 20, top 30 or top 50 to mitigate this. Others are refreshing rankings less frequently.

Google also posted (and quickly closed) a new "Engineering Analyst, Search, Anti-scraper" job role explicitly aimed at detecting and blocking SERP scraping. This poignantly underlines the retirement of the num= parameter, confirming that it was not a one-off tweak but part of a broader anti-scraping push.

My Take: Google didn’t just roll out a spam update... it rewired the data exhaust that our industry runs on at the same time. Why did they do it? Ostensibly, to minimise costs from bot activity; but if we're being completely honest here, more likely because ChatGPT and Perplexity were both recently caught scraping Google's results. Think about that for a moment – Google's biggest existential threats for over two decades were using Google to improve their competitive offerings. I'm not surprised they quickly nipped that in the bud.

Either way, expect to see quirks in rankings/volatility graphs, odd GSC deltas (impressions down, CTR and/or average position up), and some combination of higher prices, less data and slower refresh cycles from SEO tools that scrape Google results.

Operationally, a move to reporting only top 20 or top 30 rankings still gives you the core detail on what's working for you right now (guiding short-term thinking) and what's in striking distance (for the medium-term strategy). Not seeing rankings in lower positions does mean you lose some granularity that might otherwise have guided your longer-term SEO strategy. Our rank tracking tool, NeuralHQ, still tracks the top 100 positions for keywords, so for those readers working with us already, you do still have that consistency... for now.

The industry panic around this is palpable. But is it really a game-changer for contemporary SEO? I'm not so sure it is. It just means you need marketers, not data scientists, back at the helm of your long-term SEO plans. Good SEO boils down to good brand building. It's persona-based, value-driven, topically-optimised, user satisfaction-focused marketing and communications. Data helps, yes. The overarching strategy, though, should remain the same.

Of course it's super frustrating when period-in-time baselines get recalibrated and rearchitected, like we've seen in Search Console. It forces us to rethink our "go-to" reporting stack. In the end, though, this reminds me of Google removing 98% of keyword data from Google Analytics back in 2011... after all the fuss and furore, the marketers, strategists and innovators kept doing what they were doing, the results kept coming, and new ways of reporting were developed. Get the right marketers in the driving seat now and these changes will only serve to thin out the herd.

FastSearch, RankEmbed & how Google grounds AI responses

Newly surfaced documents from the Google antitrust case Remedies Memorandum Opinion (Sept 2, 2025) explain how AI Overviews are grounded (i.e. how it finds citations) via 'FastSearch'. This system pulls a much smaller set of documents much faster than the full Search algorithm, but with lower quality than the fully ranked results. A key ingredient of FastSearch is a machine-learning algorithm called RankEmbed, which leans heavily on click-and-query data (combining real Google search logs with human rater scores) to select the best contextually-relevant pages to match a user's query, based on semantic relevance, then refined for quality and intent.

My Take: There is a great deal of speculation that, as a means of saving time, FastSearch does not look at backlinks in the same way as Google's full search system. This is merely an assertion based on reading between the lines – it doesn't outright say this anywhere in the court case documents. At best, one could claim that link signals are conspicuously missing from descriptions of FastSearch. Personally, I highly doubt Google would leave PageRank out of any signal in 2025. They probably just use a much more rough-and-ready link-based signal, ignoring more advanced factors which would otherwise be used to modify and refine that signal for full search.

Either way, even if we assume it's true, focusing on that fact—“backlinks don’t matter for AIO citations”—misses the point. It misses the mechanics. If FastSearch starts from user-interaction-shaped subsets of the index, then the best way into that shortlist is to win real search demand in the first place, and links remain one of the most scalable ways to do that. That's not a coincidence, nor is it intellectual acrobatics... it's the entire reason why FastSearch works as a great proxy for the full ranking system! It's why Google do it this way.

Also, the goal isn't necessarily to get your own website cited... you want to get your brand cited! Earning backlinks from sites which get cited is a great way of getting your brand into the actual AI answer, which is the ultimate endgame for this kind of SEO (i.e. generative engine optimisation).

Of course UX is important thanks to the pattern-matching element of RankEmbed and other Google systems; but trying to optimise for pattern-matched UX signals (at least beyond best practices) is like chasing shadows. By contrast, link building is a scalable, measurable, pragmatic way to directly influence visibility.

AI Overviews, AI Mode, ChatGPT & Perplexity brand citations diverge

Fresh BrightEdge research shows increasing divergence between citations in Google AI Overviews/AI Mode and ChatGPT for the same queries. Different systems, different retrieval stacks, different brand exposure. Our own study found a ~52% overlap between AI Overview citations and the traditional organic top-20 results for the same keyword, and an uncorrelated ~46% overlap against AI Mode, due to the grounding processes being based on different query sets (the 'fanout queries' made behind-the-scenes during grounding and citation selection). It is becoming more and more apparent that there is no linear or single route to "AI visibility".

My Take: Treat “AI search” as multi-channel. One content object will not win everywhere. Create answer units mapped to distinct intents (e.g., explainer vs buyer’s guide vs troubleshooting), and target multiple retrieval modes: classic SEO for organic, citation-friendly passages for SERP Features and AI citations, entity markup for answer engines, and brand-forward, PR-style materials for LLM seeding where publisher authority is favoured. Think in terms of personas and market your brand appropriately.

Apple Intelligence: 2026 Gemini partnership?

Apple is reportedly in talks to license Google’s Gemini to power a revamped Apple Intelligence, with some upgrades publicly delayed to 2026. While nothing is signed, multiple mainstream reports (and earlier Reuters coverage) point the same way.

My Take: If Apple puts Google’s model in the loop, then it could be distribution, not model quality, that wins the game. In that scenario, you can expect Gemini-flavoured answers inside the iOS surface areas where a lot of commercial search begins. For marketers, this would mean even more weight on brand/entity optimisation and structured product data that travels well across assistants (i.e. Siri). Keep a close eye on how Siri selects products and cites sources. If it mirrors Google’s grounding biases, then AI Overview/AI Mode playbooks will port almost 1:1.

ChatGPT’s Instant Checkout ushers in agentic LLM-commerce

OpenAI has added Instant Checkout to ChatGPT – a native, in-chat purchase flow powered by its new Agentic Commerce Protocol (ACP). The first wave is US-only, starting with Etsy sellers and “over a million” Shopify merchants “coming soon”, with payments initially via Stripe. OpenAI’s docs confirm an open protocol and a lightweight merchant integration path, with access currently limited to approved partners.

OpenAI’s help materials also outline how merchant lists are ranked: by availability, price, quality, whether the seller is the maker/primary seller... and whether Instant Checkout is enabled. In short, enabling the ACP flow can itself improve your placement inside ChatGPT’s shopping UI because ChatGPT will favour it. That aside, it looks very similar to the approaches taken by Amazon, Ebay, Etsy, and perhaps unsurprisingly Shopify (Shop) when it comes to ranking products.

My Take: This is the first credible zero-click retail surface built on an LLM. The near-term playbook is commercial, not “SEO”:

- Be buyable inside ChatGPT: implement the ACP/Instant Checkout spec and ensure catalogue, price and stock data are clean and real-time. Early adopters gain distribution and a ranking nudge within ChatGPT’s merchant selector.

- Structure for retrieval: product pages need tight entities, rich product schema, and clean variants so ChatGPT can resolve items and map them to an Instant Checkout flow.

- Rethink attribution: expect leakage from traditional PPC/PLA funnels as discovery and purchase compress into ChatGPT. Start tagging ACP orders distinctly and model assist value from ChatGPT exposure.

This is assistant-led commerce becoming real. If you sell D2C, pilot the integration now; if you’re a marketplace seller, prepare for a world where the AI chat surface is the shop window!

In Other News:

1. In the US antitrust case against Google, Judge Amit Mehta imposed behavioural remedies: Google must end exclusivity deals and share certain data with rivals. But, crucially, no divestiture of Chrome/Android was forced. By contrast the EU Commission fined Google €2.95B ($3.45B) over ad-tech market abuses and signalled that divestiture remains on the table if Google can’t resolve its conflicts of interest. Political heat then rose when President Trump threatened tariff retaliation against the EU, prompting a late procedural pause and continued uncertainty.

2. Google has tested AI Overview layouts that blend with Knowledge Panels and sometimes drop the “AI Overview” label, shifting citations into columnar cards. Traditional SERPs, in certain areas, started to look indistinguishable from feature-rich AI Mode SERPs.

3. UX tests gave even more prime real estate to AI Mode. Certain keywords were spotted as showing AI Mode suggestions in the autocomplete dropdown. Clicking them took users directly to the AI Mode tab, completely bypassing traditional search. By nudging users into AI flows before the SERP even loads, Google's UX is further embracing front-of-funnel zero-click experiences.

‍