Consumers can tell the difference. 79% say AI voices should come from real, attributed actors, and 61% say a voice is most memorable when it's exclusively tied to one brand. 

Voices is the only platform that designs, licenses, and captures your brand's AI voice from real, consenting professional talent—never from scraped data. From strategic voice casting and brand calibration to consent-based rights management and studio-grade capture, we handle the entire process end to end. 

BMW, SuperBloom, and Cresta already trust Voices to build their signature AI voice. Get a demo and see how a licensed, brand-owned voice becomes a real asset—not a legal liability.

Get a Demo

On 26 August Google finished rolling out a change that got little attention outside SEO and GEO circles.

Every result link now runs through a google.com/goto redirect rather than pointing at the destination. You can still find out where a link goes, but only by following the redirect one link at a time, and a single results page carries hundreds of them. Reading Google's results used to be free. Now every link is a request you have to pay for, and at scale that adds up fast. It is the second squeeze in a year, after the num=100 parameter disappeared.

SerpApi is the obvious target, but there is something it is feeding. When you ask ChatGPT a question it does not run a single search. It fires off a batch of them in the background, reads what comes back and writes its answer from that. Those are query fan-outs, and for shopping questions a lot of them land on Google. Every one is Google's data being used to answer somebody else's customer, for free, inside a rival product.

What has been happening in the courtroom alongside it is worth a moment, because it explains the direction. Google sued SerpApi to stop it scraping and reselling search results, and in July a federal judge threw out its DMCA claims, finding Google had not shown its protections were operating with the authority of the copyright owner. Google has since gone back with an amended complaint. The technical measures, meanwhile, are achieving what the lawsuit so far has not.

On its own that reads as a Silicon Valley behemoth looking to crush a tiny vendor. Put it next to what ChatGPT's shopping answers have been doing and it looks like something a little bigger.

At Obsero we capture the underlying response rather than just the output screen, so I have the same shopping prompt stored from June and again from this week, in both the UK and the US. This allows me to put the two side by side instead of relying on a feeling that something has changed.

In June each answer came back with five products, every one carrying a written description and a price. This week the same prompt returns four, the descriptions have gone, the price stubs have gone.

That is the shopping experience getting measurably worse, in both markets, in two months.

Look at where the US links come from and the tagging gives it away. They carry utm_medium=feed, and one of them utm_campaign=openai_catalog. OpenAI is labelling those links, in the link itself, as having come out of a merchant product feed rather than off the open web.

That programme is United States only. OpenAI's own merchant page says shopping is currently live in the US and will “expand to additional regions over time, with a self-service merchant portal planned for later this year”.

So when the open supply thins out, the American answer still has a catalogue sitting underneath it. Although even that is showing wear. Of the four American links I tested this week three went through to live product pages and one went to a 404, and the dead one was the only link carrying the catalogue tag.

This is a small sample, but it makes sense given how OpenAI has been scraping content from Google.

In June every merchant link was plain utm_source=chatgpt.com and nothing else. On 30 August the same prompt returns links tagged openai_catalog and feed, others tagged genai_shopping, some tagged chatgptads, and eight local listings carrying yelp_feed.

Those tags are OpenAI telling you how a shopper reached you. feed and openai_catalog mean the merchant catalogue put you there. genai_shopping means the open web did. chatgptads means you paid for it. yelp_feed means your local listing did the work.

Nine weeks ago these links said nothing about where they came from. Now they name their source, and there is a third-party feed wired in to fill the gaps.

There is no feed programme in the UK, so when the open supply thins out there is nothing underneath to catch it. This is what that looks like on the screen. Same prompt both times, “which leggings are the most squat proof”, asked signed out on both occasions.

20 August. The lululemon Wunder Train card opened into a proper shopping panel. Named retailers, a live price against each one, stock status, delivery windows and a Visit button beside every single one of them. That is a shopper sitting one tap away from a checkout.

30 August. Same prompt, same product. The card opens to an image, a rating, a price, and then a large blank space where the retailers used to be. No sellers, no buttons, nowhere to go.

Ten days between those two screenshots, and Google completed its goto URL rollout in the middle of them.

I think so.

Back in March, research published in Search Engine Land decoded the parameters sitting inside ChatGPT's own source code and found Google Shopping sitting behind the product carousel. More than 83% of the carousel products they tested turned up as strong matches in Google's top 40 organic shopping results, across roughly 43,000 products in 10 retail categories. Bing accounted for less than 1%.

In other words the shopping carousel was, on that evidence, largely a window onto Google Shopping.

Google is making its own results harder and more expensive to reach at scale. Over the same two months ChatGPT's shopping answers have lost a product, lost their descriptions and lost their price fields. The one market where the buying experience still holds up, the US, is the one market where OpenAI went direct to merchants and built its own catalogue, and even that catalogue, in some instances, is now serving dead links.

This might be a pure coincidence and ChatGPT decides to condense the window and overall shopping experience, but I don’t think so. Google results are essential to power ChatGPT when it conducts a live retrieval and if it can’t scrape the results, then it has a quality problem.

But the direction of travel is not really in question. Whatever caused this particular fortnight, getting product data out of the open web is becoming more expensive and more fragile, and OpenAI has two ways to respond. Work harder at the sourcing, in a market Google has just made considerably more awkward. Or go and get the data from merchants directly, which means feeds, in every country, and it means the brands who are ready on the day that opens are the ones sitting in the answer with a working button beside them.

Organic traffic to retail sites is falling, and the consolation the industry keeps offering is that AI referrals convert far better, so you lose the volume and gain the quality.

That trade only pays out if the referral actually happens. A US retailer in the feed gets the link, the click and the conversion that is meant to make up for the traffic it has lost elsewhere. A British one gets named in the answer and gets no link at all, so it takes the decline and gets nothing back for it.

If you look at one prompt on one day, you have not measured anything. There is nothing to test it against, no way of knowing whether the answer you got is typical or a fluke, and no way of telling next month whether you have actually improved or just caught it on a good morning.

The fix is to group similar prompts into a topic and run them over a stretch of time. Do that and the number stops jumping about and starts meaning something, so you can say with some confidence that you are heading in the right direction, which is a claim many marketers genuinely need to make.

What you get out of it is a number you can act on: the parts of your category you own, the parts you are losing and to whom, and whether either of those is moving. I covered how many prompts that takes back in July.

Grouping on its own is not the answer either. Bolting a hundred prompts together and reporting one percentage is about as useful as reporting a single prompt, because it tells you nothing about where you are strong and where you are being taken apart.

Here is what that looks like with Levi's, a brand we track and do not work with.

Levi's runs at 45% across all topics. On its own that figure is worthless, and you could not build a plan on it if you tried.

Split it into 13 topics and it starts talking. Best jeans overall at 92.4%, age demographics at 78.9%, cuts and styles at 78.8%. Then price and value at 27.4%, use case and occasion at 24.1%, brand stockists at 14.7%, and trends and what to buy now at 4.4%.

Now you have two lists instead of one number. The strong topics are defensible ground, and you want to know what is holding them up before somebody takes them off you. The weak ones are the brief.

Within any topic you can see which competitors are being named alongside you, how close they are getting, and which citations are putting them there.

Levi's leads every topic in that view, so read it as a scoreboard and there is nothing to worry about. Look at the size of the lead instead. Best jeans overall is 92% against AGOLDE's 46%, a 46 point cushion. Fabric and comfort is 45% against Good American's 30%, and brand recommendation and quality is 35% against Wrangler's 22%. Same brand, same month, and the gap narrows from comfortable to uncomfortable as the topics get more practical.

The challenger changes as well. AGOLDE is the nearest rival on best jeans overall and sits at 1% on fabric and comfort. Good American takes 14% on best jeans overall but 34% on body type and fit. Old Navy is second on where to buy at 35%, ahead of Madewell. None of that is a brand problem, and none of it survives being averaged into a single share of voice number. Each one is a topic with a named rival and a visible set of pages putting them there, which makes it a job rather than a concern.

This is the part that makes topics extremely useful rather than just interesting. Group your prompts by where they sit in a buying decision and the topics become stages, from the first curious question all the way through to where do I actually buy this.

Levi's is dominant at the top of that journey and thin at the bottom. Strong on what makes a good pair of jeans, weak on price, weak on occasion, weak on stockists. That is a funnel problem wearing a visibility number, and a blended figure would never have shown it to you.

And then split those topics by platform, because a single topic can behave completely differently depending on who is being asked.

Take Where to buy. On ChatGPT it runs at 89% and on Gemini 83%, and on those two numbers alone you would call it a strong topic and move on to something else. Look at the same topic on Google AI Overviews and it is 47%. On Perplexity it is 40%. Same prompts, same month, a 49 point gap depending on which assistant your customer happens to be using.

That gap matters because the assistants are not reading the same web as each other. They retrieve from different underlying sources, so a page that gets you cited in one can leave you invisible in another, and the work required to fix it is different in each. Once you can see the split you can build a strategy for the platform your customers are actually on, rather than an average that describes nobody.

Aim for 25 as a working minimum.

Below that you are reporting noise. A topic with four prompts in it will swing about from day to day and hand you a number that looks precise and is not, and if that goes into a board pack you will end up defending a movement that never happened.

You can write your prompts into a spreadsheet, run them through each assistant, log whether your brand appeared, note the competitors and the cited sources, group the rows into topics and then repeat the whole thing tomorrow. It works. It also eats two or three days a month of somebody's life, and the moment you miss a week your trend line stops being a trend line.

That is the job Obsero does, across the leading AI search platforms, every day, grouped into your topics.

If you want to see what your own category looks like split this way, the free tier will show you.

Obsero is one year old on Tuesday, and starting the company is the best decision I have made in my professional career.

A year in, we are a profitable company with happy customers across many different countries, and a pipeline that is in far better shape than I would have dared predict. The thing customers keep coming back to is our ‘partner over platform’ approach, and the space itself never stops changing, which means it never stops asking you to be better than you were the day before.

How I got here matters, because I’ve worked with many great people along the way. I started out at Pi Datametrics, a leading SEO tooling company, where I met Jon Earnshaw, now a friend and a mentor, who helped shape my view of SEO and is hands down the best presenter I’ve ever seen.

I always felt I had an entrepreneurial streak in me somewhere but in my 20s/30s never seriously thought about taking the plunge and building a product of my own. You learn an enormous amount working alongside people like Jon, and it set me up for what came next at Starcom, leading on the Procter and Gamble account for four years.

If I am honest though, the most joy I have ever had in a job was at Cazoo. I got a blank canvas and the chance to build something out of absolutely nothing. We completely outperformed from an SEO point of view, beating the likes of Auto Trader, Motorway and the national and local dealers on the terms that genuinely mattered: “car finance”, “Audi A3 for sale”, “part exchange”, “sell my car”.

All of that experience, the good bits and the bad ones, is what set me up for Obsero.

Trust me, it is not a walk in the park. You are doing it with a family, with Be Franc running alongside it, in a very competitive space where a lot of the companies you are up against are carrying a great deal more VC money than you are.

Obsero started as an idea I had back in May last year. I floated it with Mike Logue, now my co-founder, and he jumped at the chance, which he covered articulately here.

If you are thinking about starting something, you need support, you need expertise, and you need somebody who will ground you and hold you accountable, and you have to do exactly the same for them.

When you find someone whose strengths fill in your gaps that completely, you have a genuinely strong proposition. Mike is fantastic and I would never have been able to do Obsero without him.

There is a lot coming that will interest this audience. An affiliate programme, an education academy and many more tools to help brands grow in the era of AI search. I will share properly over the next couple of weeks, and hopefully some very exciting news about Obsero itself.

Watch this space.

And thank you for reading. I try to document the week as I go, what the space is doing and what it is actually like to be a founder growing a business inside it, ups and downs included. That part is the best bit of doing this.

That is it for this week, until next time.

Andy