The Briefing
1. The IAB finally gave us a scorecard for AI search. On 3 August it published "Measuring Visibility in the AI Era", a 36 page standard built on four metrics: Presence, Prominence, Portrayal, Persuasion. The same document notes more than 20 companies now sell AI visibility tools and they do not agree with each other. If a vendor is pitching you, ask which of the four they actually measure.
2. Google Ads changes how your targets behave on 17 August. Budget limited campaigns on Target CPA or Target ROAS will start delivering toward your target instead of beating it. If your target is $10 and you have been getting $5, plan for $10. It is automatic, not opt in, and Google has said it will not adjust your targets for you. The Bid Target Adjustment Tool has been live since 6 July.
3. Meta shipped a first party Ads MCP server. It sits at mcp.facebook.com/ads and it is read and write, so an AI agent can pull reporting and manage campaigns through one sanctioned endpoint. Marketing API v26.0 landed on 29 July with auto upgrade and new ad account webhooks that replace polling. Your own accounts are standard access. Touching a client's account needs Advanced Access on the new ads_mcp_management permission.
4. Advantage+ now rewrites the text baked into your ad images. Spotted live on 27 July. If your creative carries burned in headlines, either fill out the Branding section under Identity (logo, fonts, colours, tone, restricted words) so Meta rewrites inside your rules, or opt out per creative. A rewritten headline that contradicts your landing page is a brand new way to break a funnel. Check your ad previews this week.
5. Yelp data is now inside ChatGPT. Reviews, ratings, photos and business details feed local answers, and the integration has since added reservations and waitlist booking without leaving the chat. If you run anything local, your reviews are no longer just a Google asset. They are training the answer a buyer sees.
The Build
My search playbook is fifteen years of habit written down. Keyword research. User intent. Technical health, on page, backlinks, internal linking, local. Everything I would tell a client to do, and none of it wrong.
Then I read the IAB document and saw the hole.
Every one of those tactics assumes a results page with ten blue links and a human choosing one. Not one of them answers the question my clients keep asking me: when someone asks ChatGPT about my category, does it mention me, and what does it say?
So I built the missing half this week. An AI Visibility Audit that runs on the IAB's four Ps.
Three parts. First, a query set: the 30 questions a real buyer types before they buy, written in their language, not keyword language. Second, running those queries across four assistants. Third, logging every answer against Presence, Prominence, Portrayal and Persuasion.
Here is the prompt that generates the query set. This is the part most people skip, and it is the part that decides whether the whole audit is worth anything.
You are building an AI Visibility Audit query set for my business. BUSINESS: [what you sell, to whom, where] CATEGORY: [the category a buyer would name] COMPETITORS: [3 to 5 names] Write 30 questions a real buyer would type into ChatGPT in the 90 days before they buy. Not keywords. Full sentences, the way a tired person types at 11pm. Split them into four groups: 1. PROBLEM AWARE (10) - they know the pain, not the fix 2. SOLUTION AWARE (10) - comparing approaches 3. VENDOR AWARE (5) - "best X for Y", "alternatives to [competitor]" 4. BRAND AWARE (5) - questions naming my business directly Rules: - No question may contain my brand name except in group 4. - Use the words a buyer uses, not the words I use. - Include at least 5 questions with a location or business-size qualifier. Output as a numbered table: question | group | what a good answer would have to contain for me to win it.
Then I ran each question and scored the answer on four lines. Mentioned or not. Where in the answer, and how substantively. Whether what was said was accurate. And whether it recommended, described, or just listed.
What worked: Portrayal was the whole ballgame. Presence is the vanity metric. Getting mentioned feels good and tells you almost nothing. Portrayal asks whether the description is accurate and whether it is the positioning you actually paid to own. That is where the useful, uncomfortable findings live. The other win was the query set itself. Buyer language beats keyword language here, because nobody types keywords into a chat window. They type sentences. Feeding voice of customer phrasing into the generator produced questions that no keyword tool would ever surface.
What didn't work: Two things. I asked the model to score its own answers against the four Ps. Bad idea. It marked itself generously and I had to move scoring to a separate pass with a clean context. Second, and bigger: these answers are not stable. Ask the same question twice and you can get two different answers with two different sources. One run is an anecdote, not a measurement. I now run every question at least three times and only count a finding if it shows up in the majority. The IAB makes the same distinction and calls it directional versus decision grade. Most of what people are calling AI visibility data right now is directional.
The Play
If you own a business with a website and you have never checked how AI describes you, you are flying blind on a channel that is already sending you buyers. You do not need a tool and you do not need a budget. You need one hour.
Run the ten question version. Use the prompt above to generate your query set, then cut it to your ten highest intent questions. Run each one three times in one assistant. Paste the answers into a fresh chat and score them with this:
Below are AI answers to buyer questions in my category.
My business is [NAME]. My competitors are [LIST].
Score every answer on four lines. Be blunt. Do not be kind.
PRESENCE: Was [NAME] mentioned? Yes or no.
PROMINENCE: If yes, first, middle, or an afterthought?
Described in detail or just listed?
PORTRAYAL: Is every claim made about [NAME] accurate?
Flag anything false. Is the framing positive,
neutral, or negative?
PERSUASION: Did the answer recommend [NAME], merely
include it, or steer the reader elsewhere?
Then give me:
- The 3 competitors mentioned most often, and why the
answers favour them.
- Every factual error stated about [NAME], quoted.
- The single question where I am closest to winning but
currently losing.
That last output is the one to act on. The near miss question is the cheapest page you will write all quarter.
Try this today. Open ChatGPT and ask it the single question your best client would have typed the week before they hired you. Run it three times. Write down whether you appeared, and whether what it said about you was true. That one line is your baseline.
Growth Partner for founders stuck between $1M and $20M | I find the money you're sitting on and build the machine that collects it | $1B+ in client results.