Don't Get Left Behind.
Ever get to the end of your workday and wonder what happened in the business world while you were busy building your empire? We got you! Max's CEO Report is the short, easy-to-read end-of-business-day recap CEOs & Founders rely on to stay ahead.
| Zymbos Intelligence · Wednesday 16 September 2026 | ||
| ||
|
A buyer asks an assistant for a shortlist in your category, and your firm is not in the answer. That moment is now a commercial fact, not a thought experiment. This week OpenAI's advertising business reached a ~£742m (~$1bn) run rate by selling sponsored cards beside ChatGPT's recommendations, its financial services edition started citing licensed sources inside every answer, Meta put a personal agent between the buyer and your website, and Google took AI search live and spoken. The through-line is simple: being found is moving from ranking in a list to being cited in an answer. Most small firms have never checked what that answer says. This issue is the check. | ||
| ||
Business · Global Paid placement is now the only guaranteed slot in an AI answer OpenAI's advertising business has reached a ~£742m (~$1bn) annualised run rate, and sponsored cards now sit directly alongside conversational recommendations in ChatGPT, according to The State of Brand. CNBC reported the same run-rate figure at the end of August. The State of Brand's conclusion is blunt: paid placement has become the only stable way to guarantee presence in an AI-generated answer, while earned mentions still decide which firms the model names on its own. Twenty years of search trained buyers to separate a blue link from a text advert because they lived in different parts of the page. In a conversational answer they share the same frame. Ask which accountancy firm to use and the model returns three or four names; on a free account a fifth may now appear beneath them because it paid. That is one specialist publication's reading, and the market is young, but the direction is clear. Visibility inside answers has acquired a price, which means the unpaid answer has acquired a value. McGann's TakeThe market has just priced "being found". If an assistant can sell the slot next to the shortlist, an unpaid mention is an asset you either hold or do not. Treat that as the reason to inspect the unpaid answer first, not the reason to buy the card tomorrow. Read more →The State of Brand | ||
Product · US ChatGPT for Financial Services writes the professional shortlist, with citations OpenAI launched ChatGPT for Financial Services on 10 September, an industry edition of ChatGPT Work built on GPT-6 Astra with design partners Morgan Stanley and Evercore, according to CNBC. It ships with licensed data from London Stock Exchange Group (LSEG), Daloopa, Crunchbase and PitchBook, source-traceable citations, firm templates and deal-room controls, and it is aimed at the pitchbook and comparables work junior analysts used to own. Pricing is undisclosed. OpenAI's own OfficeQA Pro benchmark puts Astra at 69.9 per cent against 60.2 per cent for GPT-5.6 Sol and 62.4 per cent for Claude Fable 5.1; those are vendor figures and have not been independently checked. The part that matters outside banking is the bundle. The data licensing arrives solved, and every answer carries a citation a compliance function can trace. A vertical assistant does not open ten tabs. It draws on the inventories it trusts and names who is already in them. McGann's TakeThis is what being found looks like inside a profession. When your category gets its vertical edition, the live question is whether you are a citeable source or a name the model skips. Citations are earned before they are displayed, which makes visibility work a form of due diligence. | ||
Product · Global Meta's Muse puts a personal agent between the buyer and your page Meta announced Muse on 8 September, describing it as a secure, private personal AI agent built to help people meet goals and manage daily tasks, according to the Meta Newsroom. TechCrunch asked the follow-up that matters: will consumers trust it? Muse runs inside a dedicated virtual machine, is designed to act across apps, and is meant to send mail, book travel and improve through conversation. Neither briefing states consumer pricing. The claim is not another chatbot. It is an intermediary that proposes, then acts. That is a different discovery surface from a search box that waits for a query and from a single ChatGPT tab. If Muse, or anything like it, becomes the layer that decides which accountant, which supplier, which tool, then the shortlist is assembled before the buyer sees the page you paid to rank. The agent's first impression of your firm is the compressed description it has built from whatever it could read. McGann's TakeAgents collapse the gap between "who should I use" and "book them". Visibility inside that loop is not a ranking. It is whether the agent holds a clean, consistent record of who you are, built from pages and profiles it can already fetch. |
Model release · Global Google takes AI search live, spoken and multilingual Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on 15 September, rolling them out to the Gemini API, AI Studio, Gemini Enterprise, Search Live and Gemini Live, according to Google. The models offer near real-time visual grounding and automatic switching across 97 languages. Live API pricing is ~0.4p ($0.005) per minute of audio input and ~1.3p ($0.018) per minute of output, with partners including LiveKit, Agora, Vercel, LangChain and Pipecat. Google says the Extended Thinking variant ranks first on the Artificial Analysis speech-to-speech index at 82.6; every benchmark is Google's own and no independent test appeared inside the window. For anyone trying to be found, the product name is the point: Search Live. A prospect can now ask a follow-up aloud, show the assistant a product or ask for a local recommendation without ever seeing a results page. Voice removes the friction of asking one more question, and each extra question gives the assistant another chance to narrow the field. McGann's TakeA spoken answer has no room for a muddled description. Positioning that survives a keyword may not survive a conversation. Firms that can be summarised in one clean sentence will be recommended in one; the rest get paraphrased or dropped. | ||
Policy · UK Westminster argues about brakes while the answer makers ask for rules The Cabinet Office said the UK "cannot simply turn AI off" and that blocking UK access would not stop models being built or misused elsewhere, BBC News reported on 11 September, a rejection of the legal kill switch proposed by peers and MPs. Two days later Parliament's Joint Committee on Human Rights (JCHR) called for a dedicated AI bill, a single statutory oversight body and outright prohibition of some uses; the government has two months to respond. Anthropic co-founder Jack Clark told the BBC an externally verifiable kill switch may still need to be mandatory. Meanwhile OpenAI's Tom Duff Gordon urged ministers to pass narrow, frontier-only legislation covering compulsory third-party model testing and incident reporting, POLITICO Europe reported. UK firms have operated without binding AI statute since the 2024 manifesto pledge. The live argument is who governs what an assistant is allowed to say, and who is accountable when the answer is wrong. McGann's TakeA bill will not put your firm into an answer. It may eventually tell the labs how their answers must be tested, and that raises the value of being easy to verify. Run the visibility check under today's rules. Do not wait for the statute. Read more →BBC News · BBC News on the JCHR report · BBC News, Jack Clark interview · POLITICO Europe | ||
| ||
This Week's Analysis The audit you have never run Most small firms have never once asked ChatGPT, Gemini or Claude a buyer-style question about their own category. That omission used to be harmless. This week OpenAI began selling placement beside conversational recommendations and shipped an assistant that cites licensed sources in every answer. The assistant is no longer a research side door. It is writing the shortlist. The claim, plainly: the check costs nothing, takes an afternoon, and changes behaviour immediately, because what you find is almost never what you expected. Three reasons it cannot wait. The shortlist is already being written Assistant answers already steer shortlists. A buyer who asks an assistant which firms to consider receives an answer shaped by whoever the model trusts and, increasingly, whoever paid. Meta's Muse is built to propose and then act. Google's Search Live answers a spoken question with a spoken recommendation. In each case the buyer does not open ten tabs. They accept an answer. If your firm is absent from it, you were not outranked. You were never in the room. The levers are unglamorous Citation behaviour is influenceable. Consistent naming across your website, Companies House record, LinkedIn page and any directory an assistant can fetch. One canonical description of what you sell, written in the words a buyer would type. Sources a model can quote rather than paraphrase. A firm that calls itself "consulting", "digital transformation" and "AI governance" in three disconnected places gives the model three weak signals and no stable proposition. None of this needs a retainer. It is the same hygiene that used to sit under "make the About page make sense". "Being found is no longer a position on a list. It is a sentence in an answer that a buyer may never check." A habit, not a project Once a quarter, ask the same three buyer questions of the same three assistants, paste the answers into a grid, and mark each one cited, mentioned, missing or wrong. Fix the three worst gaps. Ninety days later, run it again and measure the movement. The firm that treats this as a large programme will still be scoping the work when a competitor has already corrected a wrong category label. The strongest counter is familiar, and it is half right. Answer-engine visibility is still downstream of the same fundamentals as search: clear pages, real expertise, sources other people already trust. Chasing a separate "AI search" workstream while the website is a mess is premature optimisation. Agreed. But that is an argument against buying a toolkit before you have asked the questions yourself. It is not an argument against asking. The fundamentals overlap, which is precisely why the audit is cheap: if yours are sound, an afternoon confirms it; if they are not, you learn where the assistants lost you, which is information a rankings report never showed you. Search hands the buyer a menu they can inspect. An assistant collapses the menu into an answer, so an omission is invisible and a confident error passes unchallenged. Do the free check this week. Ask each major assistant who they would shortlist in your category, in your city, at your price point. Write down the names they give. If yours is missing, the next move is naming, sources and a single page an assistant can cite. If yours is present but wrong, the next move is a correction trail: same name, same description, same facts, everywhere a crawler already reads. Either result beats another quarter of guessing. |
| |||||||||||||||||
Semrush AI Toolkit AI visibility · Subscription · Browser
What it is Semrush is the search platform many marketing teams already pay for. Its AI Visibility Toolkit adds a dashboard that tracks how ChatGPT, Google AI, Gemini and Perplexity mention your brand: which prompts surface you, which competitors appear instead, how the answers describe you, and which pages get cited. It sits beside the position tracking and site audits Semrush users already run, so there is no separate specialist vendor to onboard. Where it helps this week Once the free manual audit in this issue has shown you a problem worth watching, Semrush turns the quarterly check into a standing report. Daily prompt tracking means you see a wrong description or a missing citation without rerunning the questions by hand, and you can watch whether a naming fix changed anything. Watch out The entry price is per domain and billed annually, so it is a commitment, not a trial you forget about. Claude and Copilot are not on the tracked list at the Base tier. The recommendations lean towards strategy rather than the operational fix. And it is only worth paying for after you have asked the assistants yourself; a monitoring tool cannot tell you whether the problem is naming, sources or absence. Ratings
Verdict Sufficient to instrument a quarterly habit. Not a substitute for asking the assistants yourself. Pricing verified on semrush.com/pricing/ai and semrush.com/pricing, 16 Sep 2026. AI Visibility Toolkit ~£73 ($99) per domain per month, billed annually; no monthly-billing price displayed. Seven-day free trial, no free plan. Additional users from ~£33 ($45) per month; Base Report ~£7 ($10) and Pro Report ~£15 ($20) per month; enterprise pricing on request. Semrush One Starter bundle (SEO plus AI Search) ~£148 ($199) per month, or ~£123 ($165.17) per month billed annually. Semrush publishes USD only; GBP converted at 1.3474. | |||||||||||||||||
| |||||||||||||||||
The Visibility Audit Audit prompt · Quarterly · Works in Claude or ChatGPT Most firms treat assistant answers as a curiosity. Treat them as a quarterly audit. This prompt turns three buyer-style questions into a grid you can reuse across ChatGPT, Claude, Gemini and Perplexity. Use it when you launch a service, change your positioning, or once a quarter as hygiene. Run it in one assistant first; it will hand you the exact questions to paste into the others, then wait. Paste the raw answers back, unedited. Wrong answers are the point: you are collecting evidence of how the assistants currently read your firm, and the same grid is what you rerun in 90 days to prove the fixes worked. Good output is a table with one row per assistant per question, a verdict of cited, mentioned, missing or wrong, the exact sentence each assistant used about you, the likely source it leaned on, and three fixes ranked by effort. Do not argue with the model. Capture what it said. You are my AI-visibility auditor. I will give you my firm name, category, market and one-sentence offer. Produce a reusable audit I can run across ChatGPT, Claude, Gemini and Perplexity. Firm (exact trading name): [NAME] Also known as: [abbreviations, former names, or "none"] Category, in the words a buyer would use: [e.g. chartered accountancy for owner-managed businesses] Market: [e.g. Manchester, UK] Offer in one sentence: [SENTENCE] Website: [URL] Three buyer questions a prospect would actually type: 1. [e.g. Who should I shortlist for X in Y?] 2. [e.g. What does NAME do, and who is it for?] 3. [e.g. NAME vs COMPETITOR: which is the better fit for Z?] Step 1. If I have not yet pasted live answers, give me the three questions formatted to paste into each assistant, tell me to run them in a fresh chat, and wait. Step 2. When I paste the answers, return a Markdown table with columns: Assistant | Question | Names cited | My firm: cited / mentioned / missing / wrong | Exact sentence about my firm, or the substitute named | Sources the answer leaned on | Priority fix: naming / description / citeable page / other. Step 3. List the three fixes that do the most work. For each: what to change, where it must appear, and how I will know it worked on the next quarterly rerun. Rules: do not invent mentions I did not paste. Label every inference as INFERRED. If the assistants disagree, show the disagreement rather than resolving it. The single fix that does the most work is consistent naming everywhere the assistants already read: your site, Companies House, LinkedIn, directories, and any page a crawler can quote. |
| ||
Closing Perspective Build the habit before someone invoices you for it The paid card inside the answer is the tell. Once a lab can sell presence next to a recommendation, visibility stops being a side effect of good pages and becomes a market. Most small firms have still never asked the assistants the question their buyers are already asking. That is the cheap problem. The expensive one arrives when an agency starts billing you to monitor a surface you have never inspected yourself. Semrush is already selling the instrumentation. OpenAI is already selling the inventory. The only open question is whether you build the habit before your agency invoices you for it. My prediction, and it is falsifiable: by 30 June 2027, AI-answer visibility reporting will be a standard priced line in new marketing agency retainers for UK professional services firms. Not a slide in a quarterly deck. A named line item, with a tool behind it and a monthly cadence. Test it by pulling 20 agency proposals that month; if fewer than ten carry the line, I was wrong. Ask ChatGPT, Gemini and Claude the same buyer question about your category this week. Reply and tell me the exact sentence each one used about your firm, or the names they offered instead. John McGann Founder, Zymbos AI |
Zymbos Intelligence zymbos.ai You're receiving this because you subscribed at zymbos.ai © 2026 Zymbos Intelligence · John McGann · London, UK Zymbos Ltd · Company No. 16198848 · Teddington, England |